Reviewed 2026-10-03. The map covers 76 publications, grouped into 74 research works, with 92 curated connections (57 documented shared mechanisms and 35 conceptual parallels). 71 works have full-text review; 3 remain partial. 2 edition pairs share stations.

Explore the homepage map · Overcomplete keyword index · Editing guide

Evidence and scope

• All 76 canonical publication entries are represented by 74 research-work stations; two conference/journal edition pairs share a station.
• 71 works have a full-text source review; HoliMol, DND, and STGG remain partial reviews and are visibly marked.
• arXiv links from the publications page are preferred, with pinned reviewed versions. Missing-page arXiv sources were identified for scTrilemma and odd-cycle matching; three further works use publisher proceedings.
• No direct real-robot sim-to-real result was established in the reviewed corpus. Simulation/proxy-to-reality is an interpretive evaluation-gap theme.
• Specific connections describe documented shared mechanisms or explicitly interpretive parallels, not automatically citations or historical influence.

The overview uses one shared station per work in a vertical figure, newest first, at every browser width. Papers are ordered by the earlier of their first arXiv version listing Sungsoo Ahn and their conference acceptance announcement. Journal publication months are a fallback when neither date is available. Spacing reserves room for labels and descriptions and is not a proportional calendar scale. The layout retains the cleaned cross-domain station placement and routes sequential connections with vertical runs and 45-degree transitions. Routing penalizes bends, detours, crossings, and close parallel tracks. Separate entry and exit positions keep domain tracks apart at shared stations. Background casings make clear gaps at unavoidable crossings, without line jumps. At widths of 992 pixels and above, lightly domain-colored boxes show descriptions and domain names beside each paper nickname. Thin grey leaders connect the boxes to their stations. Below 992 pixels, the boxes are hidden and each circle has a nearby nickname and conference/year. A stable domain legend appears below the narrow figure, sorted by first appearance.

Graph and language modeling are represented as methodologies rather than overview domains. Their application papers are assigned to molecules, proteins, materials, or cells; generic neural graph learning, reasoning, control, and optimization are grouped under General ML. Graphical models covers the classical BP, partition-function, gauge, and elimination work. Rechecking graph experiments added molecular evaluation memberships for EPIC (Section 4.1 and Table 2) and Wavelet diffusion (QM9, Sections 4.1/4.4), and biological evaluation memberships for EPIC, Non-backtracking GNN (Peptides-func/struct), and Node diffusion (PPI). Every assignment retains its source locator.

The complete map uses ordinary page scrolling at every width. Resizing changes only the label treatment, retaining the same stations, connections, and figure height. Full titles, author lists, venues, and original abstracts appear on hover. Long abstracts scroll inside the card. Papers with multiple domain memberships act as interchanges where their colored routes meet. Major lines connect only adjacent publications within each complete theme sequence and never bridge an intermediate paper.

Paper stations link directly to arXiv in a new tab. Bibliographic arXiv links take precedence; annotated arXiv sources cover papers absent from the bibliography. Existing publisher, conference, or reviewed project sources provide destinations when no arXiv record is available. Selected shared ideas appear as thin, dashed grey lines behind the solid domain routes, including before hovering. Only relationships with a specific map_idea are eligible, and their publication dates must be within the configured five-year window. Hovering a paper highlights its incident connections while its card shows only the title, authors, venue, and abstract. Idea-line hovers explain the parallel and link to its two papers. Domain-line hovers compare the papers’ actual contributions. All curated connections and their evidence remain in this analysis and the editable annotations.

All paper stations use circles; their colors indicate domains. Desktop boxes show their domain names directly; the domain legend on narrow screens follows first appearance. The analysis retains contribution families: Symmetry, Generative modeling, Representations, Learning & inference, Search & optimization, Reasoning, Agents, and Benchmarks. Contribution annotations identify the central novelty rather than every method used. QHFlow and QHFlow2 share a Symmetry contribution and an Electronic structure domain, alongside GPWNO. MaskGXT / HACO lists Agents as its main contribution and Generative modeling as a secondary contribution. Other method annotations remain available in the analysis.

Chronology selects the earlier of the first arXiv version listing Sungsoo Ahn and the official conference acceptance notification. Candidate dates and excluded pre-authorship versions are recorded below; journal months are a fallback when neither date is available. Venue/year labels remain bibliographic. Titles and identifiers break ties on the same date. Curated pairwise connections require evidence from both endpoints. The keyword pool is intentionally overcomplete and editable; keyword overlap never creates a scientific connection automatically.

Cross-paper insights

Search can train the model that replaces it

Genetic expert refinement, retrosynthesis self-improvement, planner imitation, latent veracity inference, and SGDS repeatedly turn search outcomes into reusable learned behavior. This is a useful cross-domain pattern; reward maximization, route discovery, step labeling, and distributional sampling retain different objectives.

Related works: Genetic expert-guided learning, Self-improved retrosynthesis, Planner-guided imitation, Latent veracity inference, Search-guided diffusion samplers.

Feedback inside a trajectory is a recurring bottleneck

LED-GFN explicitly decomposes terminal energy; reasoning confidence and latent veracity provide intermediate diagnostics; co-folding embeddings supply representation-level molecular feedback. Treat these as local-feedback connections, with credit assignment reserved for the stronger GFlowNet mechanism.

Related works: LED-GFN, CORE-PO, Latent veracity inference, Co-folding representations.

The choice of variables often matters as much as the learner

Gauge transformations, sparse graph grammars, rigid/flexible/all-atom MOF parameterizations, DNA segmentation, and protein token views each change what is easy to express or compute. Their constraints and guarantees differ, so factorization is a connecting question rather than one universal method.

Related works: Gauged variational inference, Gap-encoded edge lists, MOFFlow, MOFFlow-2, AtomMOF, DNAChunker, TriProRep.

Exploration has both distributional and optimization meanings

Genetic search, goal-space grids, local-search GFlowNets, adaptive teachers, and MCMC-assisted diffusion all expose insufficiently visited regions. GFlowNets and energy samplers seek distributions; genetic and combinatorial optimization principally seek good candidates.

Related works: Genetic expert-guided learning, Adaptive-grid exploration, Local Search GFlowNets, Adaptive Teachers, Search-guided diffusion samplers.

Learned fields connect several scientific domains

Electronic density, adsorption equilibrium, and spherical weather/CMB signals all use spatial representations and scale-aware approximations. Their governing physics and targets differ; learned/classical hybrids give a second connection through solver initialization or correction.

Related works: Gaussian plane-wave neural operator, MADField, Spherical neural fields, QHFlow, MCMC + belief propagation.

Scientific performance depends on what the metric measures

Protein workflow diagnostics, co-folding representation tests, Hamiltonian-to-force evaluation, crystal generation/ranking protocols, and controlled RL solver budgets explicitly separate intermediate quality from downstream utility. Proxy, simulation, and experimental validation should remain separate evidence levels.

Related works: VibeProteinBench, Co-folding representations, QHFlow2, Packora, RL4CO.

Useful invariance requires selective information retention

Failure-based debiasing and committees identify misleading easy patterns; multi-bias learning balances conflicting groups; scTrilemma separates nuisance invariance from identity and expression fidelity. These connect through information selection, while biological nuisance and classifier bias are distinct phenomena.

Related works: Learning from Failure, Learning with a biased committee, Multi-bias robust learning, scTrilemma.

Scientific language models need structure and external grounding

SMILES parsing, structural sketches, chemistry-tool agents, individualized debate, and progressive biological reasoning provide complementary routes to domain grounding. A successful tool check or confident mechanism is useful evidence, not a substitute for experimental validation.

Related works: CleanMol, Molecular Structural Reasoning, MT-Mol, INDIBATOR, PBio-Agent / LincsQA.

Paper analyses

AtomMOF

AtomMOF: All-Atom Flow Matching for MOF-Adsorbate Structure Prediction

NeurIPS 2026 · Review: fulltext.

Predicts MOF and adsorbate structures directly from the graphs of their building blocks.

Contribution. An all-atom flow model removes rigid-block constraints and uses learned interatomic potentials for Feynman-Kac steering.

Map contribution. Generative modeling: Predicts MOF and adsorbate structures directly from the graphs of their building blocks.

Evaluation. BW structural recovery and ODAC25 adsorption configurations, with model-scaling and steering studies.

Evidence boundary. Structural and adsorption-energy evaluations use computational reference datasets; generated candidates are not experimentally validated MOFs.

Other contributions. Search & optimization.

Chronology. 2026-02-07: arxiv version.

Candidate dates. 2026-02-07: arXiv 2602.07351v1; 2026-09-24: NeurIPS conference notification.

Domains. Materials (central); Molecules & drug discovery (central).

Methodologies. Flow matching (used); Equivariant models (used).

Concepts. All-atom modeling (central); Adsorption & host–guest modeling (central); Energy / reward guidance (central); Amortized inference / generation (used); Scaling studies (central).

Keywords. adsorbate; Adsorption & host–guest modeling; All-atom modeling; Amortized inference / generation; carbon capture; Diffusion Transformer; Energy / reward guidance; Equivariant models; Feynman–Kac steering; Flow matching; interatomic potential; Materials & electronic structure; MOF; Molecules & drug discovery; Scaling studies.

Sources. arXiv 2602.07351v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• MOFFlow-2 — Relax structural assumptions in MOF generation (documented); see r078.
• MADField — Host–guest modeling at different levels (documented); see r079.
• CatFlow — Joint host and adsorbate structure generation (documented); see r083.
• Packora — Flexible all-atom structure generation (documented); see r089.

Co-folding representations

A Systematic Evaluation of Co-folding Model Representations for Small-Molecule Learning

NeurIPS 2026 · Review: fulltext.

Tests whether protein–ligand co-folding can supply useful representations for standalone small molecules.

Contribution. Probing and distillation transfer Boltz2 ligand representations to property prediction, generation, and reward-based optimization.

Map contribution. Representations: The study tests and transfers ligand representations from protein–ligand co-folding models.

Evaluation. ADMET prediction, molecular generation, structure-guided optimization, and representation-alignment experiments.

Evidence boundary. Transfer is demonstrated for the evaluated model and tasks; representation-level guidance is a conceptual relative of local credit, not the LED-GFN objective.

Other contributions. Benchmarks.

Chronology. 2026-02-02: arxiv version.

Candidate dates. 2026-02-02: arXiv 2602.13249v1; 2026-09-24: NeurIPS conference notification.

Domains. Molecules & drug discovery (central); Proteins & genomics (central).

Methodologies. Representation learning (used); Reinforcement learning (used); Benchmarks & evaluation (used).

Concepts. Cross-modal learning (central); Knowledge distillation (central); Representation alignment (central); Credit assignment (conceptual parallel); Benchmarking (central).

Keywords. ADMET; Benchmarking; Benchmarks & evaluation; Boltz2; Credit assignment; Cross-modal learning; Knowledge distillation; Molecules & drug discovery; property prediction; protein–ligand co-folding; Proteins & genomics; Reinforcement learning; Representation alignment; Representation learning; reward-based optimization.

Sources. arXiv 2602.13249v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Variational information distillation — Representation transfer (documented); see r010.
• HoliMol — Molecular representations from complementary views (documented); see r047.
• LED-GFN — Dense feedback for molecular optimization (interpretive); see r052.
• TriProRep — Probe what structural representations retain (documented); see r065.
• VibeProteinBench — Evaluate scientific representations across tasks (documented); see r068.

TriProRep

Atom-level Protein Representation Learning Improves Protein Structure Prediction

NeurIPS 2026 · Review: fulltext.

Learns protein representations that support structure prediction rather than only function annotation.

Contribution. TriProRep pretrains across amino-acid, backbone, and full-atom token views; RepSP tests structural uses of the resulting embeddings.

Map contribution. Representations: Aligned sequence and atomic token views support protein structure representations.

Evaluation. Homodimer co-folding, residue interaction prediction, representation-aligned monomer prediction, and conventional function benchmarks.

Evidence boundary. The full-atom view is locally tokenized and residue-aligned; this does not imply direct generation of every atom in a whole protein.

Other contributions. Benchmarks.

Chronology. 2026-05-21: arxiv version.

Candidate dates. 2026-05-21: arXiv 2605.22133v1; 2026-09-24: NeurIPS conference notification.

Domains. Proteins & genomics (central).

Methodologies. Representation learning (used); Language / sequence models (used); Benchmarks & evaluation (used).

Concepts. All-atom modeling (central); Structured tokenization (central); Cross-modal learning (central); Representation alignment (central); Benchmarking (central).

Keywords. All-atom modeling; Benchmarking; Benchmarks & evaluation; Cross-modal learning; full-atom token; homodimer; Language / sequence models; Proteins & genomics; Representation alignment; Representation learning; RepSP; residue interaction; structural probing; Structured tokenization.

Sources. arXiv 2605.22133v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• DNAChunker — Biological tokenization as a modeling decision (documented); see r064.
• Co-folding representations — Probe what structural representations retain (documented); see r065.
• Antibody sequence–structure decoupling — Sequence and structure need compatible representations (documented); see r066.

scTrilemma

scTrilemma: Balancing Identity, Invariance, and Reconstruction in Single-Cell Representation Learning

NeurIPS 2026 · Review: fulltext.

Balances biological identity, nuisance invariance, and expression fidelity in single-cell embeddings.

Contribution. A latent-bottleneck VAE routes information through the embedding, decoder, and context-conditioned prior under one reconstruction objective.

Map contribution. Representations: The model routes biological and nuisance information through distinct parts of a representation.

Evaluation. Release-based zero-shot evaluation, disease cohorts, differential-expression analyses, and latent interventions.

Evidence boundary. The paper’s source title now says Fidelity while the canonical publication title says Reconstruction; neither title is silently rewritten here.

Other contributions. Generative modeling.

Chronology. 2026-09-24: conference notification.

Candidate dates. 2026-09-24: NeurIPS conference notification; 2026-09-30: arXiv 2609.38840v1.

Domains. Cells & perturbations (central).

Methodologies. Representation learning (used); Probabilistic inference (used); Robust learning (used).

Concepts. Information routing (central); Invariance & spurious factors (central); Competing objectives (central); Benchmarking (used).

Keywords. batch effects; Benchmarking; Cells & perturbations; Competing objectives; context-conditioned prior; Information routing; Invariance & spurious factors; latent bottleneck; Probabilistic inference; Representation learning; Robust learning; single-cell RNA; variational autoencoder.

Sources. arXiv 2609.38840v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Variational information distillation — What information should an embedding retain? (interpretive); see r011.
• Multi-bias robust learning — Invariance has competing costs (interpretive); see r016.

Packora

Packora: Systematic Design for Generative Molecular Crystal Structure Prediction

arXiv 2026 · Review: fulltext.

Generates molecular crystal structures while separating candidate generation from downstream ranking.

Contribution. A flexible all-atom flow model supports multiple components and optional structural conditions, studied under controlled generation and ranking protocols.

Map contribution. Generative modeling: Generates molecular crystal structures while separating candidate generation from downstream ranking.

Evaluation. Six generation benchmarks and common relaxation/ranking pipelines inspired by crystal-structure blind tests.

Evidence boundary. Experimental-form recovery measures agreement with known crystal forms; it is not evidence of prospective synthesis or a general robotics sim-to-real result.

Other contributions. Benchmarks.

Chronology. 2026-08-27: arxiv version.

Candidate dates. 2026-08-27: arXiv 2608.26962v1.

Domains. Materials (central); Molecules & drug discovery (central).

Methodologies. Flow matching (used); Equivariant models (used); Benchmarks & evaluation (used).

Concepts. All-atom modeling (central); Conditional generation (central); Benchmarking (central); Simulation / proxy-to-reality gap (evaluation); Scaling studies (central).

Keywords. All-atom modeling; Benchmarking; Benchmarks & evaluation; blind test; Conditional generation; Equivariant models; Flow matching; Materials & electronic structure; molecular crystal; Molecules & drug discovery; polymorph; ranking; relaxation; Scaling studies; Simulation / proxy-to-reality gap.

Sources. arXiv 2608.26962v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• VibeProteinBench — Generation quality depends on the evaluation pipeline (interpretive); see r069.
• Multimodal Crystal Flow — Conditional periodic generative models (documented); see r085.
• AtomMOF — Flexible all-atom structure generation (documented); see r089.
• QHFlow2 — Evaluate the downstream quantity, control the pipeline (interpretive); see r092.

MaskGXT / HACO

Discovering Crystal Structure Prediction Algorithms with an AI Co-Scientist

arXiv 2026 · Review: fulltext.

An AI co-scientist develops masked generative models for crystal structure prediction.

Contribution. HACO searches methodological families and refines MaskGXT with symmetry tokens, polymorph coverage, and coordinate refinement.

Map contribution. Agents: An AI co-scientist searches and refines crystal-prediction algorithms with expert guidance.

Evaluation. MP-20 and MPTS-52 structure recovery plus polymorph-coverage and component ablations.

Evidence boundary. The algorithm-discovery demonstration assumes cheap, aligned validation; results do not establish a generally autonomous scientific discovery system.

Other contributions. Generative modeling.

Chronology. 2026-06-22: arxiv version.

Candidate dates. 2026-06-22: arXiv 2606.22866v1.

Domains. Materials (central).

Methodologies. Language / sequence models (used); Search & optimization (used).

Concepts. Structured tokenization (central); Symmetry (central); Diversity & mode coverage (central); Method transfer (central); Algorithm discovery (central); Benchmarking (used).

Keywords. Algorithm discovery; Benchmarking; Diversity & mode coverage; HACO; Language & scientific reasoning; Language / sequence models; masked generative modeling; MaskGXT; Materials & electronic structure; Method transfer; polymorph coverage; Search & optimization; Structured tokenization; Symmetry; symmetry token.

Sources. arXiv 2606.22866v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• CleanMol — Structural tokens expose hidden scientific constraints (interpretive); see r071.
• Multimodal Crystal Flow — Composition and symmetry structure crystal generation (documented); see r087.
• Genetic expert-guided learning — A specialized search partner improves a generator (interpretive); see r088.

MADField

MADField: Multi-fidelity Amortized Density Field for Adsorption in Nanoporous Materials

arXiv 2026 · Review: fulltext.

Estimates adsorption equilibrium as a spatial density field instead of repeatedly simulating individual gas particles.

Contribution. Multi-fidelity cDFT and GCMC supervision supports uptake integration, screening, and initialization of a classical solver.

Map contribution. Learning & inference: Multi-fidelity training amortizes adsorption-equilibrium prediction into a density-field model.

Evaluation. Density and uptake prediction, transfer to disordered hosts, ARC-MOF screening, and cDFT initialization.

Evidence boundary. Higher fidelity here means GCMC relative to cDFT under the studied simulation setting, not direct experimental ground truth.

Chronology. 2026-06-19: arxiv version.

Candidate dates. 2026-06-19: arXiv 2606.21284v1.

Domains. Materials (central).

Methodologies. Neural fields & operators (used).

Concepts. Density / field representations (central); Adsorption & host–guest modeling (central); Multi-fidelity learning (central); Amortized inference / generation (central); Learned / classical hybrids (central); Simulation / proxy-to-reality gap (conceptual parallel).

Keywords. Adsorption & host–guest modeling; Amortized inference / generation; ARC-MOF; cDFT; Density / field representations; gas uptake; GCMC; Learned / classical hybrids; Materials & electronic structure; Multi-fidelity learning; Neural fields & operators; Simulation / proxy-to-reality gap; solver initialization.

Sources. arXiv 2606.21284v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• AtomMOF — Host–guest modeling at different levels (documented); see r079.
• Gaussian plane-wave neural operator — A physical field replaces repeated microscopic computation (interpretive); see r080.
• MCMC + belief propagation — A fast approximation benefits from another solver (interpretive); see r082.

VibeProteinBench

VibeProteinBench: An Evaluation Benchmark for Language-interfaced Vibe Protein Design

arXiv 2026 · Review: fulltext.

Evaluates protein recognition, engineering, and generation through natural-language requests.

Contribution. Expert-curated rationales and computational checks expose distinct failure modes in language-interfaced protein design.

Map contribution. Benchmarks: The benchmark separates recognition, engineering, and generation failures in protein-design workflows.

Evaluation. General and specialized language models, with and without tools, across three design-workflow stages.

Evidence boundary. In silico plausibility checks do not demonstrate experimental protein function or successful wet-lab design.

Chronology. 2026-05-09: arxiv version.

Candidate dates. 2026-05-09: arXiv 2605.10978v1.

Domains. Proteins & genomics (central).

Methodologies. Language / sequence models (used); Benchmarks & evaluation (used).

Concepts. Benchmarking (central); Tool-grounded reasoning (central); Simulation / proxy-to-reality gap (evaluation); Cross-modal learning (used).

Keywords. Benchmarking; Benchmarks & evaluation; Cross-modal learning; expert rationale; in silico validation; Language & scientific reasoning; Language / sequence models; natural-language design; protein engineering; protein recognition; Proteins & genomics; Simulation / proxy-to-reality gap; Tool-grounded reasoning.

Sources. arXiv 2605.10978v4: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Antibody sequence–structure decoupling — Diversity does not establish biological function (interpretive); see r067.
• Co-folding representations — Evaluate scientific representations across tasks (documented); see r068.
• Packora — Generation quality depends on the evaluation pipeline (interpretive); see r069.

AdaPert

Learning Adaptive Perturbation-Conditioned Contexts for Robust Transcriptional Response Prediction

ICML 2026 · Review: fulltext.

Predicts perturbation-specific transcriptional responses without collapsing to an average cell response.

Contribution. A differentiable, sparse knowledge-graph context and adaptive signal–noise treatment isolate responsive genes.

Map contribution. Learning & inference: Predicts perturbation-specific transcriptional responses without collapsing to an average cell response.

Evaluation. Four CRISPR perturbation datasets, DEG-aware metrics, and cross-cell-line analyses.

Evidence boundary. Generalization is tested on held-out perturbations and cell lines; causal intervention claims require additional evidence.

Other contributions. Representations.

Chronology. 2026-02-21: arxiv version.

Candidate dates. 2026-02-21: arXiv 2602.18885v1; 2026-04-30: ICML conference notification.

Domains. Cells & perturbations (central).

Methodologies. Graph neural networks (used); Robust learning (used).

Concepts. Context selection (central); Signal–noise separation (central); Graph structure & transformations (central); Benchmarking (used).

Keywords. Benchmarking; cell-line transfer; Cells & perturbations; Context selection; CRISPR; differential expression; Graph neural networks; Graph structure & transformations; knowledge graph; perturbation response; Robust learning; Signal–noise separation.

Sources. arXiv 2602.18885v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Multi-bias robust learning — Separate signal from misleading regularity (interpretive); see r022.
• PBio-Agent / LincsQA — Perturbation-specific biological reasoning (documented); see r023.

INDIBATOR

INDIBATOR: Diverse and Fact-Grounded Individuality for Multi-Agent Debate in Molecular Discovery

arXiv 2026 · Review: fulltext.

Grounds molecular-discovery debates in individualized scientific and molecular histories.

Contribution. Agents propose, critique, and vote using fine-grained profiles with literature knowledge and structural priors.

Map contribution. Agents: Agents propose, critique, and select molecules using individual scientific histories.

Evaluation. Protein-conditioned generation, bioactivity-guided generation, and goal-directed lead optimization.

Evidence boundary. Profile diversity and benchmark performance do not establish that every agent’s scientific claim is correct or every molecule works experimentally.

Other contributions. Reasoning.

Chronology. 2026-02-02: arxiv version.

Candidate dates. 2026-02-02: arXiv 2602.01815v1.

Domains. Molecules & drug discovery (central).

Methodologies. Language / sequence models (used); Search & optimization (used).

Concepts. Diversity & mode coverage (central); Tool-grounded reasoning (central); Multi-agent deliberation (central).

Keywords. agent profiles; bioactivity; Diversity & mode coverage; Language & scientific reasoning; Language / sequence models; lead optimization; Molecules & drug discovery; Multi-agent deliberation; scientific debate; Search & optimization; Tool-grounded reasoning; voting.

Sources. arXiv 2602.01815v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Disentangled risk-sensitive MARL — Multiple agents contribute different uncertainty or viewpoints (interpretive); see r036.
• MT-Mol — Deliberation for molecule discovery (documented); see r073.
• PBio-Agent / LincsQA — Ground scientific agents in domain context (documented); see r074.

Multimodal Crystal Flow

Multimodal Crystal Flow: Any-to-Any Modality Generation for Unified Crystal Modeling

ICML 2026 · Review: fulltext.

Unifies crystal generation tasks that previously required task-specific models.

Contribution. Separate flow times for atom types and geometry allow any-to-any conditioning, with composition-aware ordering and symmetry augmentation.

Map contribution. Generative modeling: Unifies crystal generation tasks that previously required task-specific models.

Evaluation. CSP, de novo generation, and structure-conditioned atom-type generation on MP-20 and MPTS-52.

Evidence boundary. A shared generative model does not remove the need to validate stability and the consistency of requested conditions.

Chronology. 2026-02-23: arxiv version.

Candidate dates. 2026-02-23: arXiv 2602.20210v1; 2026-04-30: ICML conference notification.

Domains. Materials (central).

Methodologies. Flow matching (used).

Concepts. Cross-modal learning (central); Conditional generation (central); Symmetry (central).

Keywords. any-to-any generation; composition ordering; Conditional generation; Cross-modal learning; crystal structure prediction; Flow matching; Materials & electronic structure; separate flow times; Symmetry.

Sources. arXiv 2602.20210v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Packora — Conditional periodic generative models (documented); see r085.
• MOFFlow-2 — Discrete chemistry meets continuous geometry (documented); see r086.
• MaskGXT / HACO — Composition and symmetry structure crystal generation (documented); see r087.

PBio-Agent / LincsQA

Progressive Multi-Agent Reasoning for Biological Perturbation Prediction

arXiv 2026 · Review: fulltext.

Predicts gene-regulation directions under chemical and genetic perturbations with structured biological reasoning.

Contribution. LincsQA tests chemical mechanisms; PBio-Agent sequences tasks by difficulty and feeds confident predictions into harder cases.

Map contribution. Agents: The agent sequences biological tasks and passes confident results to harder predictions.

Evaluation. LincsQA compound perturbations and PerturbQA genetic perturbations, with judge-based verification.

Evidence boundary. The outputs are regulation directions and explanations, not a complete quantitative simulator of a perturbed cell.

Other contributions. Reasoning.

Chronology. 2026-02-07: arxiv version.

Candidate dates. 2026-02-07: arXiv 2602.07408v1.

Domains. Cells & perturbations (central).

Methodologies. Language / sequence models (used); Benchmarks & evaluation (used).

Concepts. Tool-grounded reasoning (central); Multi-agent deliberation (central); Curriculum & training difficulty (central); Benchmarking (central); Context selection (central).

Keywords. Benchmarking; Benchmarks & evaluation; Cells & perturbations; Context selection; Curriculum & training difficulty; gene regulation; Language & scientific reasoning; Language / sequence models; LincsQA; mechanism of action; Multi-agent deliberation; PerturbQA; progressive prediction; Tool-grounded reasoning.

Sources. arXiv 2602.07408v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• AdaPert — Perturbation-specific biological reasoning (documented); see r023.
• Causal Influence Prompting — Explicit reasoning context before a decision (interpretive); see r024.
• INDIBATOR — Ground scientific agents in domain context (documented); see r074.

Riemannian MeanFlow

Riemannian MeanFlow

ICML 2026 · Review: fulltext.

Learns generative flow maps directly on manifolds for rapid biological design.

Contribution. Equivalent average-velocity identities support one- and few-step generation and inexpensive terminal-state reward look-ahead.

Map contribution. Generative modeling: Learns generative flow maps directly on manifolds for rapid biological design.

Evaluation. Promoter DNA design and protein backbone generation, with parameterization and guided-inference studies.

Evidence boundary. A few-step flow map and an equilibrium sampler solve different training problems; this work is evaluated on the reported biological design tasks.

Other contributions. Search & optimization.

Chronology. 2026-02-08: arxiv version.

Candidate dates. 2026-02-08: arXiv 2602.07744v1; 2026-04-30: ICML conference notification.

Domains. Proteins & genomics (central); General ML (central).

Methodologies. Flow matching (used); Equivariant models (used).

Concepts. Manifold geometry (central); Few-step generation (central); Energy / reward guidance (central).

Keywords. average velocity; Energy / reward guidance; Equivariant models; Few-step generation; flow map; Flow matching; General learning & inference; Manifold geometry; promoter DNA; protein backbone; Proteins & genomics; terminal look-ahead.

Sources. arXiv 2602.07744v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• BioEmu-CV — Reuse a generative model for biological computation (interpretive); see r059.
• MOFFlow — Generation on non-Euclidean variables (documented); see r060.

QHFlow2

Machine Learning Hamiltonians are Accurate Energy-Force Predictors

ICML 2026 · Review: fulltext.

Tests predicted electronic Hamiltonians by the energies and forces they actually produce.

Contribution. QHFlow2 combines an SO(2) backbone and two-stage pair updates under a common downstream evaluation pipeline.

Map contribution. Symmetry: The SO(2)-equivariant backbone and pair updates structure Hamiltonian prediction.

Evaluation. MD17/rMD17 and QH9 Hamiltonian, energy, and force evaluations with capacity and data scaling.

Evidence boundary. The QHFlow2 name alone is not evidence that the model uses the flow-matching objective of QHFlow; Hamiltonian prediction is treated separately here.

Other contributions. Learning & inference; Benchmarks.

Chronology. 2026-02-18: arxiv version.

Candidate dates. 2026-02-18: arXiv 2602.16897v1; 2026-04-30: ICML conference notification.

Domains. Electronic structure (central).

Methodologies. Graph neural networks (used); Equivariant models (used); Benchmarks & evaluation (used).

Concepts. Symmetry (central); Benchmarking (central); Scaling studies (central); Physical observables (central).

Keywords. Benchmarking; Benchmarks & evaluation; electronic Hamiltonian; energy and forces; Equivariant models; Graph neural networks; Materials & electronic structure; Molecules & drug discovery; Physical observables; QH9; Scaling studies; SO(2); Symmetry; two-stage pair update.

Sources. arXiv 2602.16897v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• QHFlow — Hamiltonian prediction versus downstream physics (documented); see r090.
• Packora — Evaluate the downstream quantity, control the pipeline (interpretive); see r092.

CatFlow

CatFlow: Co-generation of Slab-Adsorbate Systems via Flow Matching

ICML 2026 · Review: fulltext.

Co-generates catalyst surfaces and adsorbate configurations as a coupled structure.

Contribution. A primitive-cell factorization reduces redundant variables and preserves slab–adsorbate orientation in a flow model.

Map contribution. Generative modeling: A joint generative model uses primitive-cell variables for catalyst surfaces and adsorbates.

Evaluation. De novo and conditional slab–adsorbate generation with structural and thermodynamic comparisons.

Evidence boundary. OC20 structure and adsorption-energy evaluations assess computational interfaces rather than measured catalyst performance.

Other contributions. Representations.

Chronology. 2026-02-05: arxiv version.

Candidate dates. 2026-02-05: arXiv 2602.05372v1; 2026-04-30: ICML conference notification.

Domains. Materials (central); Molecules & drug discovery (central).

Methodologies. Flow matching (used); Equivariant models (used).

Concepts. Adsorption & host–guest modeling (central); Factorization & decomposition (central); Conditional generation (central); Physics-informed priors (central).

Keywords. Adsorption & host–guest modeling; catalyst; Conditional generation; Equivariant models; Factorization & decomposition; Flow matching; Materials & electronic structure; Molecules & drug discovery; OC20; orientation; Physics-informed priors; primitive cell; slab–adsorbate.

Sources. arXiv 2602.05372v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• AtomMOF — Joint host and adsorbate structure generation (documented); see r083.
• MOFFlow — Factorization reduces periodic structural complexity (documented); see r084.

Latent veracity inference

Latent Veracity Inference for Identifying Errors in Stepwise Reasoning

ICLR 2026 · Review: fulltext.

Infers which individual steps in a reasoning chain are incorrect.

Contribution. Veracity Search explores latent correctness assignments and supplies pseudo-labels for an amortized veracity model.

Map contribution. Reasoning: Latent correctness estimates identify errors in individual reasoning steps.

Evaluation. Logical, mathematical, and commonsense reasoning, plus self-correction and self-improvement applications.

Evidence boundary. Language-model likelihood is a proxy reward; accurate step identification on benchmarks is not a universal truth guarantee.

Other contributions. Search & optimization; Learning & inference.

Chronology. 2025-05-17: arxiv version.

Candidate dates. 2025-05-17: arXiv 2505.11824v1; 2026-01-25: ICLR conference notification.

Domains. General ML (central).

Methodologies. Probabilistic inference (used); Language / sequence models (used); Search & optimization (used).

Concepts. Amortized inference / generation (central); Search–learning feedback (central); Local feedback (central); Verification & diagnostics (central).

Keywords. Amortized inference / generation; General learning & inference; Language & scientific reasoning; Language / sequence models; Local feedback; Probabilistic inference; pseudo-label; reasoning chain; Search & optimization; Search–learning feedback; self-correction; step correctness; Veracity Search; Verification & diagnostics.

Sources. arXiv 2505.11824v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• LED-GFN — Locate feedback inside a long construction (interpretive); see r053.
• CORE-PO — Reasoning paths matter beyond final answers (documented); see r075.
• Self-improved retrosynthesis — Successful search supplies reusable supervision (interpretive); see r076.

MELD

Learning Flexible Forward Trajectories for Masked Molecular Diffusion

ICLR 2026 · Review: fulltext.

Addresses collisions between masked-diffusion trajectories of distinct molecular graphs.

Contribution. A learned element-wise corruption schedule keeps partially corrupted graphs distinguishable during reconstruction.

Map contribution. Generative modeling: Learned corruption paths keep different molecular graphs distinguishable during generation.

Evaluation. Unconditional and property-conditioned molecular generation on QM9, ZINC250K, GuacaMol, and polymer tasks.

Evidence boundary. Reported chemical validity is a dataset-specific generation result, not evidence of synthesizability or biological utility.

Chronology. 2025-05-22: arxiv version.

Candidate dates. 2025-05-22: arXiv 2505.16790v1; 2026-01-25: ICLR conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. Diffusion models (used); Graph neural networks (used).

Concepts. Adaptive trajectories (central); Structured tokenization (central); Constraint-aware design (central).

Keywords. Adaptive trajectories; Constraint-aware design; corruption schedule; Diffusion models; Graph neural networks; masked diffusion; molecular validity; Molecules & drug discovery; polymer; Structured tokenization; trajectory collision.

Sources. arXiv 2505.16790v4: Abstract; method formulation; experimental results and limitations.

Specific connections.

• EPIC — Choose informative transformation paths (interpretive); see r019.

BioEmu-CV

Learning Collective Variables from BioEmu with Time-Lagged Generation

ICLR 2026 · Review: fulltext.

Learns slow collective variables from time-lagged conditioning of a pretrained protein ensemble generator.

Contribution. A frozen BioEmu backbone is adapted through an encoder that retains slow dynamical information for enhanced sampling.

Map contribution. Representations: The encoder extracts slow dynamical information from a pretrained protein generator.

Evaluation. Free-energy estimation, CV-steered transition paths, and interpretability on fast-folding proteins.

Evidence boundary. Generated equilibrium samples and time-lagged training do not by themselves identify exact transition kinetics.

Chronology. 2025-07-10: arxiv version.

Candidate dates. 2025-07-10: arXiv 2507.07390v1; 2026-01-25: ICLR conference notification.

Domains. Proteins & genomics (central).

Methodologies. Representation learning (used); Diffusion models (used).

Concepts. Rare events & transition paths (central); Time scales & dynamics (central); Learned / classical hybrids (central); Conditional generation (central); Benchmarking (central).

Keywords. Benchmarking; BioEmu; collective variables; Conditional generation; Diffusion models; enhanced sampling; free energy; Learned / classical hybrids; Proteins & genomics; Rare events & transition paths; Representation learning; time-lagged conditioning; Time scales & dynamics.

Sources. arXiv 2507.07390v4: Abstract; method formulation; experimental results and limitations.

Specific connections.

• TPS-DPS — Slow dynamics and rare protein transitions (documented); see r058.
• Riemannian MeanFlow — Reuse a generative model for biological computation (interpretive); see r059.

DNAChunker

DNACHUNKER: Learnable Tokenization for DNA Language Models

ICML 2026 · Review: fulltext.

Learns the segmentation of DNA jointly with a masked language model.

Contribution. Context-dependent variable-length units allocate detail to functionally enriched regions and compress redundant sequence.

Map contribution. Representations: Learned genomic units allocate representation detail according to sequence context.

Evaluation. Five genomic benchmarks, mutation resilience, segmentation analyses, and component ablations.

Evidence boundary. Adaptive genomic tokenization does not assume that DNA has fixed natural-language word boundaries.

Chronology. 2026-01-06: arxiv version.

Candidate dates. 2026-01-06: arXiv 2601.03019v1; 2026-04-30: ICML conference notification.

Domains. Proteins & genomics (central).

Methodologies. Language / sequence models (used); Representation learning (used).

Concepts. Structured tokenization (central); Compression (central); Adaptive resolution (central).

Keywords. Adaptive resolution; Compression; genomic segmentation; Language / sequence models; masked language model; mutation resilience; Proteins & genomics; Representation learning; Structured tokenization; variable-length tokens.

Sources. arXiv 2601.03019v4: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Layer-adaptive pruning — Adaptive allocation of representational budget (interpretive); see r012.
• Gap-encoded edge lists — Choose tokens that respect the object (interpretive); see r044.
• TriProRep — Biological tokenization as a modeling decision (documented); see r064.

RL4CO

RL4CO: an Extensive Reinforcement Learning for Combinatorial Optimization Benchmark

KDD 2025 · Review: fulltext.

Provides a common framework for comparing reinforcement-learning solvers for combinatorial problems.

Contribution. Modular environments, policies, algorithms, and controlled studies lower implementation and evaluation overhead.

Map contribution. Benchmarks: A common framework supports controlled comparisons of combinatorial reinforcement-learning solvers.

Evaluation. Routing solver comparisons with matched architectures and training-sample budgets, alongside broad problem coverage.

Evidence boundary. Library coverage and benchmark findings depend on the reviewed version and controlled settings; current library popularity is not used as evidence.

Other contributions. Search & optimization.

Chronology. 2024-06-21: arxiv version. The map uses the first arXiv version listing Sungsoo Ahn as a coauthor. Earlier versions without him are excluded. KDD 2025 has November 2024 and May 2025 notification rounds; both follow the June 2024 coauthored arXiv version. The paper-specific round is not verified.

Candidate dates. 2024-06-21: arXiv 2306.17100v4.

Excluded versions. 2306.17100v1 (2023-06-29): Sungsoo Ahn is absent from this version’s author list.; 2306.17100v2 (2023-09-13): Sungsoo Ahn is absent from this version’s author list.; 2306.17100v3 (2023-12-04): Sungsoo Ahn is absent from this version’s author list.

Domains. General ML (central).

Methodologies. Reinforcement learning (used); Benchmarks & evaluation (used).

Concepts. Benchmarking (central); Combinatorial optimization (central); Modular frameworks (central).

Keywords. Benchmarking; Benchmarks & evaluation; Combinatorial optimization; Graphs & discrete problems; matched sample budgets; Modular frameworks; Reinforcement learning; routing; solver library; traveling salesman problem.

Sources. arXiv 2306.17100v6: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Learning what to defer — Learning discrete solvers and judging them fairly (documented); see r026.
• Symmetric replay training — Sample budgets in learned combinatorial optimization (documented); see r029.

Energy-based generator matching

Energy-based Generator Matching: A Neural Sampler for General State Space

NeurIPS 2025 · Review: fulltext.

Trains neural samplers directly from energy functions on continuous, discrete, or mixed spaces.

Contribution. Generator matching uses self-normalized importance estimates and bootstrapping across flow, diffusion, and jump processes.

Map contribution. Generative modeling: Trains neural samplers directly from energy functions on continuous, discrete, or mixed spaces.

Evaluation. Discrete systems and joint discrete–continuous distributions with comparisons of bootstrapped estimators.

Evidence boundary. Empirical tests cover the reported discrete and mixed-state systems; modality-agnostic formulation does not imply unlimited practical scale.

Chronology. 2025-05-26: arxiv version.

Candidate dates. 2025-05-26: arXiv 2505.19646v1; 2025-09-18: NeurIPS conference notification.

Domains. General ML (central).

Methodologies. Energy-based sampling (used); Flow matching (used); Diffusion models (used).

Concepts. Amortized inference / generation (central); Unnormalized distributions (central); Discrete–continuous states (central); Importance sampling (central); Variance reduction (central).

Keywords. Amortized inference / generation; bootstrapping; Diffusion models; Discrete–continuous states; Energy-based sampling; Flow matching; General learning & inference; generator matching; Importance sampling; jump process; self-normalized importance sampling; Unnormalized distributions; Variance reduction.

Sources. arXiv 2505.19646v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Iterated energy-based flow matching — Learn generators from energies (documented); see r054.
• Search-guided diffusion samplers — Reusable sampling under expensive energy calls (documented); see r055.

Search-guided diffusion samplers

On Scalable and Efficient Training of Diffusion Samplers

NeurIPS 2025 · Review: fulltext.

Improves neural sampling when energy calls are expensive and landscapes contain difficult-to-reach modes.

Contribution. An MCMC Searcher with novelty rewards feeds an off-policy diffusion Learner; periodic resets address primacy bias.

Map contribution. Generative modeling: Improves neural sampling when energy calls are expensive and landscapes contain difficult-to-reach modes.

Evaluation. Standard energy benchmarks, higher-dimensional systems, and molecular conformer generation.

Evidence boundary. Mixing and learned mode coverage remain task-dependent; improved finite-budget sampling is not a universal exact-sampling claim.

Other contributions. Search & optimization.

Chronology. 2025-05-26: arxiv version.

Candidate dates. 2025-05-26: arXiv 2505.19552v1; 2025-09-18: NeurIPS conference notification.

Domains. General ML (central); Molecules & drug discovery (central).

Methodologies. Energy-based sampling (used); Diffusion models (used); Search & optimization (used).

Concepts. MCMC (central); Amortized inference / generation (central); Search–learning feedback (central); Exploration–exploitation (central); Replay & off-policy reuse (central); Diversity & mode coverage (central).

Keywords. Amortized inference / generation; conformer; diffusion Learner; Diffusion models; Diversity & mode coverage; Energy-based sampling; Exploration–exploitation; General learning & inference; MCMC; MCMC Searcher; Molecules & drug discovery; novelty reward; primacy bias; Replay & off-policy reuse; Search & optimization; Search–learning feedback.

Sources. arXiv 2505.19552v4: Abstract; method formulation; experimental results and limitations.

Specific connections.

• MCMC + belief propagation — Approximation with a stochastic search partner (interpretive); see r007.
• Energy-based generator matching — Reusable sampling under expensive energy calls (documented); see r055.
• Adaptive Teachers — Coverage-driven amortized sampling (documented); see r056.
• TPS-DPS — Off-policy molecular sampling (documented); see r057.

MOFFlow-2

Flexible MOF Generation with Torsion-Aware Flow Matching

NeurIPS 2025 · Review: fulltext.

Generates MOF building blocks and assembles flexible structures with novel chemistry.

Contribution. A two-stage model combines SMILES generation with flow matching over block translations, rotations, torsions, and lattice variables.

Map contribution. Generative modeling: Generates MOF building blocks and assembles flexible structures with novel chemistry.

Evaluation. MOF generation and structure prediction with novel-block and validity analyses.

Evidence boundary. Structure initialization and a metal library remain inputs to the pipeline; flexible blocks are not equivalent to an unconstrained all-atom model.

Chronology. 2025-05-23: arxiv version.

Candidate dates. 2025-05-23: arXiv 2505.17914v1; 2025-09-18: NeurIPS conference notification.

Domains. Materials (central); Molecules & drug discovery (central).

Methodologies. Flow matching (used); Language / sequence models (used); Equivariant models (used).

Concepts. Factorization & decomposition (central); Manifold geometry (central); Conditional generation (central); Constraint-aware design (central).

Keywords. Conditional generation; Constraint-aware design; Equivariant models; Factorization & decomposition; flexible building block; Flow matching; Language / sequence models; lattice; Manifold geometry; Materials & electronic structure; metal library; Molecules & drug discovery; SMILES; torsion.

Sources. arXiv 2505.17914v4: Abstract; method formulation; experimental results and limitations.

Specific connections.

• MOFFlow — From rigid to flexible MOF assembly (documented); see r077.
• AtomMOF — Relax structural assumptions in MOF generation (documented); see r078.
• Multimodal Crystal Flow — Discrete chemistry meets continuous geometry (documented); see r086.

QHFlow

High-order Equivariant Flow Matching for Density Functional Theory Hamiltonian Prediction

NeurIPS 2025 · Review: fulltext.

Predicts structured electronic Hamiltonians to reduce expensive self-consistent-field computations.

Contribution. Equivariant flow matching and orbital-energy alignment produce physically informed Hamiltonian estimates.

Map contribution. Symmetry: Equivariant vector fields preserve the transformation rules of electronic Hamiltonians.

Evaluation. MD17 and QH9 Hamiltonian prediction plus SCF initialization and runtime comparisons.

Evidence boundary. Hamiltonian reconstruction quality and downstream energy–force accuracy are different tests; the later QHFlow2 study investigates the latter.

Other contributions. Generative modeling.

Chronology. 2025-05-24: arxiv version.

Candidate dates. 2025-05-24: arXiv 2505.18817v1; 2025-09-18: NeurIPS conference notification.

Domains. Electronic structure (central).

Methodologies. Flow matching (used); Equivariant models (used).

Concepts. Symmetry (central); Learned / classical hybrids (central); Physical observables (central); Conditional generation (central).

Keywords. Conditional generation; electronic Hamiltonian; Equivariant models; Flow matching; Learned / classical hybrids; Materials & electronic structure; Molecules & drug discovery; orbital alignment; Physical observables; QH9; SCF initialization; Symmetry.

Sources. arXiv 2505.18817v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• QHFlow2 — Hamiltonian prediction versus downstream physics (documented); see r090.
• Gaussian plane-wave neural operator — Amortize electronic-structure computation (documented); see r091.

CORE-PO

Self-Training Large Language Models with Confident Reasoning

EMNLP 2025 · Review: fulltext.

Improves reasoning self-training by evaluating reasoning paths rather than only final answers.

Contribution. Policy optimization prefers paths with high reasoning-level confidence to avoid learning accidental correct answers.

Map contribution. Reasoning: Training uses confidence in reasoning paths instead of only final-answer correctness.

Evaluation. Four in-distribution and two out-of-distribution reasoning benchmarks.

Evidence boundary. Confidence remains a model-derived signal and can be miscalibrated; reasoning quality is evaluated on the reported tasks.

Other contributions. Search & optimization.

Chronology. 2025-05-23: arxiv version.

Candidate dates. 2025-05-23: arXiv 2505.17454v1; 2025-08-20: EMNLP conference notification.

Domains. General ML (central).

Methodologies. Language / sequence models (used); Reinforcement learning (used).

Concepts. Self-training (central); Local feedback (central); Verification & diagnostics (central).

Keywords. accidental correct answers; Language & scientific reasoning; Language / sequence models; Local feedback; policy optimization; pseudo-label; reasoning confidence; Reinforcement learning; Self-training; Verification & diagnostics.

Sources. arXiv 2505.17454v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• BootGen — Pseudo-labels shape the next learner (interpretive); see r063.
• Latent veracity inference — Reasoning paths matter beyond final answers (documented); see r075.

MT-Mol

MT-Mol: Multi Agent System with Tool-based Reasoning for Molecular Optimization

EMNLP 2025 · Review: fulltext.

Optimizes molecules with specialized agents that use chemistry tools to inspect proposals.

Contribution. Analyst, scientist, verifier, and reviewer agents combine RDKit feedback with stepwise reasoning.

Map contribution. Agents: Specialist agents use chemistry tools to check and refine molecular proposals.

Evaluation. PMO-1K molecular optimization with tool, agent, and budget comparisons.

Evidence boundary. Tool consistency and low-budget proxy scores do not prove synthetic feasibility or experimental activity.

Other contributions. Search & optimization; Reasoning.

Chronology. 2025-05-27: arxiv version.

Candidate dates. 2025-05-27: arXiv 2505.20820v1; 2025-08-20: EMNLP conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. Language / sequence models (used); Search & optimization (used).

Concepts. Tool-grounded reasoning (central); Multi-agent deliberation (central); Local feedback (central).

Keywords. chemist agents; Language & scientific reasoning; Language / sequence models; Local feedback; low-budget optimization; Molecules & drug discovery; Multi-agent deliberation; PMO-1K; RDKit; Search & optimization; Tool-grounded reasoning.

Sources. arXiv 2505.20820v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Molecular Structural Reasoning — Intermediate molecular reasoning with checks (documented); see r072.
• INDIBATOR — Deliberation for molecule discovery (documented); see r073.

CleanMol

Improving Chemical Understanding of LLMs via SMILES Parsing

EMNLP 2025 · Review: fulltext.

Teaches language models to recover molecular graph information from SMILES strings.

Contribution. Deterministic subgraph and global-graph parsing tasks provide scalable, difficulty-aware structural supervision.

Map contribution. Representations: Graph-parsing supervision teaches language models to represent molecular connectivity.

Evaluation. SMILES parsing diagnostics and transfer to Mol-Instructions tasks.

Evidence boundary. Parsing accuracy addresses a prerequisite for molecular reasoning, rather than proving complete chemical understanding.

Chronology. 2025-05-22: arxiv version.

Candidate dates. 2025-05-22: arXiv 2505.16340v1; 2025-08-20: EMNLP conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. Language / sequence models (used); Representation learning (used).

Concepts. Graph structure & transformations (central); Structured tokenization (central); Curriculum & training difficulty (central); Verification & diagnostics (central).

Keywords. Curriculum & training difficulty; global graph; Graph structure & transformations; Language & scientific reasoning; Language / sequence models; Mol-Instructions; Molecules & drug discovery; Representation learning; SMILES parsing; Structured tokenization; subgraph parsing; Verification & diagnostics.

Sources. arXiv 2505.16340v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Molecular Structural Reasoning — Recover molecular structure from text (documented); see r070.
• MaskGXT / HACO — Structural tokens expose hidden scientific constraints (interpretive); see r071.

Causal Influence Prompting

Enhancing LLM Agent Safety via Causal Influence Prompting

ACL 2025 · Review: fulltext.

Uses causal influence diagrams to guide safer language-model agent decisions.

Contribution. Agents initialize, consult, and refine explicit diagrams of actions, uncertain factors, and utilities.

Map contribution. Reasoning: Explicit causal diagrams structure actions, uncertainty, and utilities during agent reasoning.

Evaluation. Code-execution and mobile-control safety, including side effects and adversarial evaluations.

Evidence boundary. Generated causal diagrams are reasoning aids rather than independently identified causal models.

Other contributions. Agents.

Chronology. 2025-05-15: conference notification.

Candidate dates. 2025-05-15: ACL conference notification; 2025-07-01: arXiv 2507.00979v1.

Domains. General ML (central).

Methodologies. Language / sequence models (used).

Concepts. Causal reasoning & interventions (central); Verification & diagnostics (central); Context selection (central).

Keywords. agent safety; causal influence diagram; Causal reasoning & interventions; Context selection; decision utilities; Language & scientific reasoning; Language / sequence models; side effects; Verification & diagnostics.

Sources. arXiv 2507.00979v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• PBio-Agent / LincsQA — Explicit reasoning context before a decision (interpretive); see r024.
• Concept intervention analysis — An interpretable edit needs a valid intervention protocol (interpretive); see r025.

Molecular Structural Reasoning

Structural Reasoning Improves Molecular Understanding of LLM

ACL 2025 · Review: fulltext.

Improves language-model molecular understanding through explicit structural sketches.

Contribution. Structure-based intermediate reasoning supports tasks with a known molecule and tasks that must generate one.

Map contribution. Reasoning: Intermediate structural sketches make molecular reasoning explicit.

Evaluation. Molecular reasoning diagnostics and fine-tuning on three molecular tasks.

Evidence boundary. Structure-grounded reasoning can reduce errors without making language models reliable chemical simulators.

Other contributions. Representations.

Chronology. 2024-10-08: arxiv version.

Candidate dates. 2024-10-08: arXiv 2410.05610v1; 2025-05-15: ACL conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. Language / sequence models (used); Representation learning (used).

Concepts. Graph structure & transformations (central); Structured tokenization (central); Local feedback (central).

Keywords. Graph structure & transformations; intermediate reasoning; Language & scientific reasoning; Language / sequence models; Local feedback; molecular graph reconstruction; Molecules & drug discovery; Representation learning; structural sketch; Structured tokenization.

Sources. arXiv 2410.05610v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• CleanMol — Recover molecular structure from text (documented); see r070.
• MT-Mol — Intermediate molecular reasoning with checks (documented); see r072.

RxnFlow

Generative Flows on Synthetic Pathway for Drug Design

ICLR 2025 · Review: fulltext.

Designs diverse molecules through explicit synthetic reaction pathways.

Contribution. GFlowNet training operates over building blocks and reaction templates, with subsampling of a large action space.

Map contribution. Search & optimization: Searches reaction pathways for diverse high-reward molecules within an available synthesis space.

Evaluation. Pocket-specific optimization, pocket-conditioned generation, and changes to objectives and building-block libraries.

Evidence boundary. Template-based synthetic feasibility and docking scores do not guarantee an experimentally executable route or active drug.

Other contributions. Generative modeling.

Chronology. 2024-10-06: arxiv version.

Candidate dates. 2024-10-06: arXiv 2410.04542v1; 2025-01-22: ICLR conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. GFlowNets (used); Search & optimization (used).

Concepts. Synthesis & reactant availability (central); Constraint-aware design (central); Diversity & mode coverage (central); Combinatorial optimization (central); Conditional generation (central).

Keywords. action subsampling; building-block library; Combinatorial optimization; Conditional generation; Constraint-aware design; Diversity & mode coverage; docking; GFlowNets; Molecules & drug discovery; pocket conditioning; reaction template; Search & optimization; Synthesis & reactant availability.

Sources. arXiv 2410.04542v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• RetCL — Availability constrains molecular design (documented); see r038.
• Local Search GFlowNets — Reward-proportional molecular search (documented); see r039.
• Spanning-tree molecular generation — Feasibility built into the action space (documented); see r040.

MOFFlow

MOFFlow: Flow Matching for Structure Prediction of Metal-Organic Frameworks

ICLR 2025 · Review: fulltext.

Predicts MOF structures by assembling metal nodes and organic linkers as rigid building blocks.

Contribution. Riemannian flow matching models blockwise rotations, translations, and lattice variables in a reduced search space.

Map contribution. Generative modeling: Predicts MOF structures by assembling metal nodes and organic linkers as rigid building blocks.

Evaluation. Structural recovery, speed, and scaling with the number of atoms and building blocks.

Evidence boundary. Rigid-block assumptions constrain local geometry; later flexible and all-atom models address this restriction.

Other contributions. Representations.

Chronology. 2024-10-07: arxiv version.

Candidate dates. 2024-10-07: arXiv 2410.17270v1; 2025-01-22: ICLR conference notification.

Domains. Materials (central).

Methodologies. Flow matching (used); Equivariant models (used).

Concepts. Factorization & decomposition (central); Manifold geometry (central); Conditional generation (central); Amortized inference / generation (used).

Keywords. Amortized inference / generation; Conditional generation; Equivariant models; Factorization & decomposition; Flow matching; lattice; Manifold geometry; Materials & electronic structure; MOF; rigid building block; rotation; translation.

Sources. arXiv 2410.17270v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Riemannian MeanFlow — Generation on non-Euclidean variables (documented); see r060.
• MOFFlow-2 — From rigid to flexible MOF assembly (documented); see r077.
• CatFlow — Factorization reduces periodic structural complexity (documented); see r084.

TPS-DPS

Transition Path Sampling with Improved Off-Policy Training of Diffusion Path Samplers

ICLR 2025 · Review: fulltext.

Samples molecular transition paths without requiring manually designed collective variables.

Contribution. An off-policy diffusion path sampler combines log-variance training, control variates, replay, annealing, and equivariant bias forces.

Map contribution. Generative modeling: Samples molecular transition paths without requiring manually designed collective variables.

Evaluation. Double-well, alanine dipeptide, and four fast-folding protein transition-path tasks.

Evidence boundary. Transition-path distributions differ from equilibrium coordinate distributions; this method does not simply replace every MD observable.

Other contributions. Symmetry.

Chronology. 2024-05-30: arxiv version.

Candidate dates. 2024-05-30: arXiv 2405.19961v1; 2025-01-22: ICLR conference notification.

Domains. Proteins & genomics (central); Molecules & drug discovery (central).

Methodologies. Diffusion models (used); Energy-based sampling (used); Equivariant models (used).

Concepts. Rare events & transition paths (central); Amortized inference / generation (central); Replay & off-policy reuse (central); Variance reduction (central); Time scales & dynamics (central).

Keywords. alanine dipeptide; Amortized inference / generation; control variates; Diffusion models; Energy-based sampling; equivariant bias; Equivariant models; log-variance loss; Molecules & drug discovery; Proteins & genomics; Rare events & transition paths; Replay & off-policy reuse; Time scales & dynamics; transition path sampling; Variance reduction.

Sources. arXiv 2405.19961v5: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Search-guided diffusion samplers — Off-policy molecular sampling (documented); see r057.
• BioEmu-CV — Slow dynamics and rare protein transitions (documented); see r058.

ReBind

ReBind: Enhancing Ground-state Molecular Conformation Prediction via Force-Based Graph Rewiring

ICLR 2025 · Review: fulltext.

Improves conformation prediction for atoms whose geometry depends strongly on non-bonded interactions.

Contribution. Force-aware graph rewiring adds interactions motivated by Lennard-Jones potentials, especially for low-degree atoms.

Map contribution. Learning & inference: Force-derived graph edges improve predictions of non-bonded interactions.

Evaluation. Ground-state conformation benchmarks across molecular sizes and analyses by atom degree.

Evidence boundary. Force-based rewiring encodes a modeling prior; it is not a full first-principles treatment of all intermolecular forces.

Other contributions. Representations.

Chronology. 2024-10-04: arxiv version.

Candidate dates. 2024-10-04: arXiv 2410.14696v1; 2025-01-22: ICLR conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. Graph neural networks (used); Equivariant models (used).

Concepts. Graph structure & transformations (central); Physics-informed priors (central); All-atom modeling (used).

Keywords. All-atom modeling; conformation; Equivariant models; Graph neural networks; graph rewiring; Graph structure & transformations; Lennard-Jones; low-degree atoms; Molecules & drug discovery; non-bonded interactions; Physics-informed priors.

Sources. arXiv 2410.14696v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Non-backtracking GNNs — Change graph communication to preserve useful interactions (interpretive); see r020.

Adaptive Teachers

Adaptive Teachers for Amortized Samplers

ICLR 2025 · Review: fulltext.

Directs amortized samplers toward states their current training policy misses.

Contribution. A learned Teacher targets high-loss regions of the Student to improve mode coverage and sample efficiency.

Map contribution. Search & optimization: Targets poorly explored regions to improve sampler coverage and search efficiency.

Evaluation. Synthetic exploration, continuous diffusion sampling, and biochemical design with teacher and local-search ablations.

Evidence boundary. High loss is a proxy for insufficient coverage; a teacher is not an external oracle or an exact posterior sampler.

Other contributions. Generative modeling.

Chronology. 2024-10-02: arxiv version.

Candidate dates. 2024-10-02: arXiv 2410.01432v1; 2025-01-22: ICLR conference notification.

Domains. General ML (central); Molecules & drug discovery (central); Proteins & genomics (central).

Methodologies. GFlowNets (used); Energy-based sampling (used).

Concepts. Exploration–exploitation (central); Diversity & mode coverage (central); Curriculum & training difficulty (central); Auxiliary teachers (central); Amortized inference / generation (central); Replay & off-policy reuse (central).

Keywords. Amortized inference / generation; Auxiliary teachers; continuous sampling; Curriculum & training difficulty; Diversity & mode coverage; Energy-based sampling; Exploration–exploitation; General learning & inference; GFlowNets; high-loss regions; mode coverage; Molecules & drug discovery; Proteins & genomics; Replay & off-policy reuse; Student–Teacher.

Sources. arXiv 2410.01432v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Learning from Failure — An auxiliary model redirects training (interpretive); see r015.
• Difficult, not too different — Informative difficulty as a curriculum (interpretive); see r018.
• Adaptive-grid exploration — Spend exploration where coverage is weak (interpretive); see r035.
• Local Search GFlowNets — Coverage through additional training trajectories (documented); see r051.
• Search-guided diffusion samplers — Coverage-driven amortized sampling (documented); see r056.

Antibody sequence–structure decoupling

Decoupled Sequence and Structure Generation for Realistic Antibody Design

TMLR 2024 · Review: fulltext.

Separates antibody sequence generation from structure prediction to improve architecture flexibility.

Contribution. Decoupling and a composition-based objective address excessive token repetition in non-autoregressive design.

Map contribution. Generative modeling: Separate sequence generation and structure prediction improve antibody-design flexibility.

Evaluation. SAbDAb, RAbD, affinity optimization, and CDR-H3 design with docked templates.

Evidence boundary. Sequence repetition and computational affinity metrics are proxies for developability, not clinical safety evidence.

Chronology. 2024-05-27: arxiv version. The map uses the first arXiv version listing Sungsoo Ahn as a coauthor. Earlier versions without him are excluded. The TMLR archive lists January 2025 while the bibliography lists 2024; the map now sorts by the earlier May 2024 coauthored arXiv version and preserves the bibliography.

Candidate dates. 2024-05-27: arXiv 2402.05982v2; 2025-01: TMLR journal month.

Excluded versions. 2402.05982v1 (2024-02-08): Sungsoo Ahn is absent from this version’s author list.

Domains. Proteins & genomics (central).

Methodologies. Representation learning (used); Equivariant models (used).

Concepts. Factorization & decomposition (central); Diversity & mode coverage (central); Constraint-aware design (central); Competing objectives (central).

Keywords. affinity proxy; antibody; CDR-H3; Competing objectives; composition objective; Constraint-aware design; Diversity & mode coverage; Equivariant models; Factorization & decomposition; Proteins & genomics; Representation learning; token repetition.

Sources. arXiv 2402.05982v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• TriProRep — Sequence and structure need compatible representations (documented); see r066.
• VibeProteinBench — Diversity does not establish biological function (interpretive); see r067.

Structurally diverse molecular LLMs

Can LLMs Generate Diverse Molecules? Towards Alignment with Structural Diversity

arXiv 2024 · Review: fulltext.

Aligns a language model with structural diversity across generated molecules.

Contribution. A set-autoregressive formulation and reinforcement learning reward chemical differences rather than only textual differences.

Map contribution. Generative modeling: Aligns a language model with structural diversity across generated molecules.

Evaluation. Description-guided molecular generation against diverse sequence-decoding baselines.

Evidence boundary. Structural diversity is one design objective and does not establish potency or broad experimental validity.

Other contributions. Search & optimization.

Chronology. 2024-10-04: arxiv version.

Candidate dates. 2024-10-04: arXiv 2410.03138v1.

Domains. Molecules & drug discovery (central).

Methodologies. Language / sequence models (used); Reinforcement learning (used).

Concepts. Diversity & mode coverage (central); Conditional generation (central); Simulation / proxy-to-reality gap (conceptual parallel).

Keywords. Conditional generation; description-guided generation; Diversity & mode coverage; Language & scientific reasoning; Language / sequence models; Molecules & drug discovery; Reinforcement learning; set-autoregressive; Simulation / proxy-to-reality gap; structural diversity.

Sources. arXiv 2410.03138v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• RoMA — A proxy improvement can miss the intended goal (interpretive); see r062.

Iterated energy-based flow matching

Iterated Energy-based Flow Matching for Sampling from Boltzmann Densities

arXiv 2024 · Review: fulltext.

Learns a continuous generator from an energy function when target samples are unavailable.

Contribution. A simulation-free, off-policy flow-matching objective estimates the marginal vector field and reuses prior samples.

Map contribution. Generative modeling: Learns a continuous generator from an energy function when target samples are unavailable.

Evaluation. Two-dimensional Gaussian mixtures and an eight-dimensional four-particle double-well distribution.

Evidence boundary. The reported experiments are low-dimensional GMM and double-well systems; larger scientific systems require additional validation.

Chronology. 2024-08-29: arxiv version.

Candidate dates. 2024-08-29: arXiv 2408.16249v1.

Domains. General ML (central).

Methodologies. Energy-based sampling (used); Flow matching (used).

Concepts. Unnormalized distributions (central); Amortized inference / generation (central); Importance sampling (central); Replay & off-policy reuse (central).

Keywords. Amortized inference / generation; double well; Energy-based sampling; Flow matching; Gaussian mixture; General learning & inference; Importance sampling; off-policy flow matching; partition function; Replay & off-policy reuse; Unnormalized distributions.

Sources. arXiv 2408.16249v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• MCMC + belief propagation — Unnormalized targets, different estimators (interpretive); see r006.
• Energy-based generator matching — Learn generators from energies (documented); see r054.

Non-backtracking GNNs

Non-backtracking Graph Neural Networks

TMLR 2024 · Review: fulltext.

Reduces redundant message flows that revisit the immediately preceding graph node.

Contribution. Non-backtracking updates improve information propagation, with theoretical analyses of sensitivity and expressivity.

Map contribution. Learning & inference: Reduces redundant message flows that revisit the immediately preceding graph node.

Evaluation. Long-range graph benchmarks, stochastic block models, and transductive node classification.

Evidence boundary. Suppressing immediate backtracking addresses a particular source of redundancy rather than all long-range reasoning limitations.

Chronology. 2023-10-11: arxiv version.

Candidate dates. 2023-10-11: arXiv 2310.07430v1; 2024-09: TMLR journal month.

Domains. General ML (central); Proteins & genomics (evaluation).

Methodologies. Graph neural networks (used).

Concepts. Message passing (central); Graph structure & transformations (central); Long-range propagation (central).

Keywords. expressivity; Graph neural networks; Graph structure & transformations; Graphs & discrete problems; Long-range propagation; Message passing; node sensitivity; non-backtracking; oversquashing; stochastic block model.

Sources. arXiv 2310.07430v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• ReBind — Change graph communication to preserve useful interactions (interpretive); see r020.
• Structured node diffusion — Beyond independent local node decisions (documented); see r021.

Pessimistic backward policy

Pessimistic Backward Policy for GFlowNets

NeurIPS 2024 · Review: fulltext.

Addresses under-exploitation of high-reward objects in GFlowNet training.

Contribution. A learned pessimistic backward policy reduces unobserved backward flow and aligns observed flow with terminal rewards.

Map contribution. Search & optimization: Improves exploitation of high-reward objects through pessimistic backward-policy training.

Evaluation. Hyper-grid, bags, structured sets, molecules, and RNA design tasks.

Evidence boundary. Pessimistic refers to a particular backward-flow optimization; it is not a general conservative offline-RL guarantee.

Other contributions. Generative modeling.

Chronology. 2024-05-25: arxiv version.

Candidate dates. 2024-05-25: arXiv 2405.16012v1; 2024-09-26: NeurIPS conference notification.

Domains. General ML (central); Molecules & drug discovery (central); Proteins & genomics (central).

Methodologies. GFlowNets (used).

Concepts. Exploration–exploitation (central); Diversity & mode coverage (central); Adaptive trajectories (central); Amortized inference / generation (used).

Keywords. Adaptive trajectories; Amortized inference / generation; backward policy; Diversity & mode coverage; Exploration–exploitation; General learning & inference; GFlowNets; Molecules & drug discovery; Proteins & genomics; RNA design; terminal rewards; unobserved backward flow.

Sources. arXiv 2405.16012v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• LED-GFN — Improve GFlowNet training without changing its target (documented); see r049.
• Local Search GFlowNets — Backward trajectories affect discovery (documented); see r050.

Spherical neural fields

Hybrid Neural Representation for Spherical Data

ICML 2024 · Review: fulltext.

Represents weather and cosmic microwave background signals on the sphere.

Contribution. Hierarchical spherical feature grids and interpolated positional features support a coordinate-based neural representation.

Map contribution. Representations: Hierarchical spherical grids represent physical signals at multiple spatial scales.

Evaluation. Regression, super-resolution, temporal interpolation, and compression on weather and CMB data.

Evidence boundary. Weather/CMB fields, adsorption densities, and electron densities share a field representation but have different governing physics.

Chronology. 2024-02-05: arxiv version.

Candidate dates. 2024-02-05: arXiv 2402.05965v1; 2024-05-01: ICML conference notification.

Domains. Weather & cosmology (central).

Methodologies. Neural fields & operators (used).

Concepts. Manifold geometry (central); Density / field representations (conceptual parallel); Adaptive resolution (central); Compression (central).

Keywords. Adaptive resolution; Compression; cosmic microwave background; Density / field representations; HNeR; Manifold geometry; Neural fields & operators; spherical feature grid; super-resolution; weather; Weather & cosmology.

Sources. arXiv 2402.05965v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Gaussian plane-wave neural operator — Represent spatial signals with mixed scales (interpretive); see r081.

Gaussian plane-wave neural operator

Gaussian Plane-Wave Neural Operator for Electron Density Estimation

ICML 2024 · Review: fulltext.

Predicts electronic density using complementary local and global basis functions.

Contribution. Gaussian orbital and plane-wave components capture high- and low-frequency structure in a neural operator.

Map contribution. Representations: Local Gaussian and global plane-wave functions represent complementary electron-density scales.

Evaluation. QM9, molecular dynamics, and Materials Project electron-density datasets.

Evidence boundary. The paper explicitly leaves downstream exchange-correlation-energy evaluation unresolved; density error is not itself an energy-accuracy guarantee.

Chronology. 2024-02-05: arxiv version.

Candidate dates. 2024-02-05: arXiv 2402.04278v1; 2024-05-01: ICML conference notification.

Domains. Electronic structure (central).

Methodologies. Neural fields & operators (used); Equivariant models (used).

Concepts. Density / field representations (central); Factorization & decomposition (central); Physical observables (central); Physics-informed priors (central).

Keywords. Density / field representations; electron density; Equivariant models; exchange-correlation; Factorization & decomposition; frequency decomposition; Gaussian orbital; Materials & electronic structure; Molecules & drug discovery; Neural fields & operators; Physical observables; Physics-informed priors; plane wave.

Sources. arXiv 2402.04278v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Wavelet graph diffusion — Separate coarse and fine spatial information (interpretive); see r046.
• HoliMol — Local and global components complement one another (interpretive); see r048.
• MADField — A physical field replaces repeated microscopic computation (interpretive); see r080.
• Spherical neural fields — Represent spatial signals with mixed scales (interpretive); see r081.
• QHFlow — Amortize electronic-structure computation (documented); see r091.

Multi-bias robust learning

Improving Robustness to Multiple Spurious Correlations by Multi-Objective Optimization

ICML 2024 · Review: fulltext.

Handles multiple spurious correlations whose mitigation objectives can conflict.

Contribution. Dynamic weighting of group losses seeks a minimax Pareto solution, with a new multi-bias benchmark.

Map contribution. Learning & inference: Handles multiple spurious correlations whose mitigation objectives can conflict.

Evaluation. MultiCelebA, UrbanCars, Multi-Color MNIST, and conventional single-bias datasets.

Evidence boundary. Training assumes bias-attribute annotations, unlike the earlier failure-based and committee-based methods.

Other contributions. Benchmarks.

Chronology. 2024-05-01: conference notification.

Candidate dates. 2024-05-01: ICML conference notification; 2024-09-05: arXiv 2409.03303v1.

Domains. General ML (central).

Methodologies. Robust learning (used).

Concepts. Competing objectives (central); Invariance & spurious factors (central); Signal–noise separation (central); Benchmarking (used).

Keywords. Benchmarking; bias attributes; Competing objectives; General learning & inference; group loss; Invariance & spurious factors; minimax Pareto; MultiCelebA; Robust learning; Signal–noise separation; UrbanCars.

Sources. arXiv 2409.03303v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Learning with a biased committee — Spurious-correlation mitigation (documented); see r014.
• scTrilemma — Invariance has competing costs (interpretive); see r016.
• AdaPert — Separate signal from misleading regularity (interpretive); see r022.

Symmetric replay training

Enhancing Sample Efficiency in Black-box Combinatorial Optimization via Symmetric Replay Training

ICML 2024 · Review: fulltext.

Makes black-box combinatorial optimization more sample-efficient when reward calls are costly.

Contribution. Replaying solution-preserving symmetric trajectories supplies extra training examples without new oracle evaluations.

Map contribution. Symmetry: Equivalent solution trajectories provide additional training examples without new reward evaluations.

Evaluation. Traveling salesman, hardware design, and molecular optimization tasks.

Evidence boundary. The useful symmetry must preserve the solution and reward; arbitrary data transformations do not provide free evaluation.

Other contributions. Search & optimization.

Chronology. 2023-06-02: arxiv version.

Candidate dates. 2023-06-02: arXiv 2306.01276v1; 2024-05-01: ICML conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. Reinforcement learning (used); Search & optimization (used).

Concepts. Symmetry (central); Replay & off-policy reuse (central); Combinatorial optimization (central); Exploration–exploitation (central); Sample efficiency (central).

Keywords. Combinatorial optimization; Exploration–exploitation; Graphs & discrete problems; hardware design; Molecules & drug discovery; oracle budget; Reinforcement learning; Replay & off-policy reuse; Sample efficiency; Search & optimization; symmetric trajectories; Symmetry; traveling salesman problem.

Sources. arXiv 2306.01276v4: Abstract; method formulation; experimental results and limitations.

Specific connections.

• RL4CO — Sample budgets in learned combinatorial optimization (documented); see r029.
• Genetic expert-guided learning — Reuse good candidates under expensive rewards (documented); see r030.

Unsupervised combinatorial optimization

Tackling Complex Conditions in Unsupervised Combinatorial Optimization

ICML 2024 · Review: fulltext.

Constructs differentiable objectives and rounding rules for constrained discrete optimization.

Contribution. Theory-guided probabilistic objectives and derandomization handle cardinality, minimum, and related common conditions.

Map contribution. Search & optimization: Constructs differentiable objectives and rounding rules for constrained discrete optimization.

Evaluation. Facility location, maximum coverage, and additional synthetic and real-world graph problems.

Evidence boundary. The theoretical properties apply to the derived objectives and conditions, not every combinatorial constraint.

Chronology. 2024-05-01: conference notification.

Candidate dates. 2024-05-01: ICML conference notification; 2024-05-14: arXiv 2405.08424v1.

Domains. General ML (central).

Methodologies. Search & optimization (used); Probabilistic inference (used).

Concepts. Combinatorial optimization (central); Constraint-aware design (central); Derandomization (central).

Keywords. cardinality; Combinatorial optimization; Constraint-aware design; Derandomization; facility location; Graphs & discrete problems; maximum coverage; Probabilistic inference; probabilistic objective; rounding; Search & optimization.

Sources. arXiv 2405.08424v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Learning what to defer — Constrained discrete decisions (documented); see r027.
• Blossom belief propagation — Theory guides the discrete solver (interpretive); see r028.

Adaptive-grid exploration

Breadth-First Exploration in Adaptive Grid-based Reinforcement Learning

ICML 2024 · Review: fulltext.

Explores long-horizon goal spaces while recording both achieved and unattained subgoals.

Contribution. Breadth-first probing on an adaptive grid prioritizes coarse coverage before costly local refinement.

Map contribution. Search & optimization: Explores long-horizon goal spaces while recording both achieved and unattained subgoals.

Evaluation. Goal-conditioned control, including constrained fixed-goal settings and exploration analyses.

Evidence boundary. The evaluated control tasks are simulated; no real-robot sim-to-real result is inferred from these experiments.

Chronology. 2024-05-01: conference notification.

Candidate dates. 2024-05-01: ICML conference notification.

Domains. General ML (central).

Methodologies. Reinforcement learning (used); Search & optimization (used).

Concepts. Exploration–exploitation (central); Adaptive resolution (central); Graph structure & transformations (central); Constraint-aware design (central).

Keywords. adaptive grid; Adaptive resolution; breadth-first exploration; Constraint-aware design; Control & multi-agent learning; Exploration–exploitation; goal-conditioned RL; Graph structure & transformations; Reinforcement learning; Search & optimization; unattained subgoal.

Sources. Conference proceedings: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Planner-guided imitation — Subgoal structure for long-horizon control (documented); see r034.
• Adaptive Teachers — Spend exploration where coverage is weak (interpretive); see r035.

HoliMol

Holistic Molecular Representation Learning via Multi-view Fragmentation

TMLR 2024 · Review: partial.

Connects global molecular graph features to chemically meaningful local fragments.

Contribution. Fragmentation-based contrastive learning uses molecular substructures and an additional 3D geometry view.

Map contribution. Representations: Fragment and geometry views support transferable molecular representations.

Evaluation. The accessible abstract reports molecular property prediction; implementation uses paired 2D/3D molecular data.

Evidence boundary. Only indexed publisher excerpts and the official author implementation were accessible; detailed experiments and limitations remain partially reviewed.

No arXiv link is listed for this work. This draft uses the accessible abstract and, where available, author implementation; full-paper review remains pending.

Chronology. 2024-06: journal month. No arXiv link or independently verifiable journal acceptance date is available in the source annotations. The verified June 2024 journal month is retained as a fallback.

Candidate dates. 2024-06: TMLR journal month.

Domains. Molecules & drug discovery (central).

Methodologies. Representation learning (used); Graph neural networks (used).

Concepts. Factorization & decomposition (central); Cross-modal learning (central).

Keywords. 2D–3D views; Cross-modal learning; Factorization & decomposition; fragment contrast; Graph neural networks; molecular fragmentation; Molecules & drug discovery; property prediction; Representation learning.

Sources. Author implementation: README and indexed abstract; partial review.

Specific connections.

• Co-folding representations — Molecular representations from complementary views (documented); see r047.
• Gaussian plane-wave neural operator — Local and global components complement one another (interpretive); see r048.

EPIC

EPIC: Graph Augmentation with Edit Path Interpolation via Learnable Cost

IJCAI 2024 · Review: fulltext.

Interpolates between graphs through meaningful edit operations for data augmentation.

Contribution. A learned context-sensitive edit cost guides paths between graphs rather than interpolating raw adjacency matrices.

Map contribution. Learning & inference: Learned graph-edit costs construct useful augmentation paths.

Evaluation. Eleven graph-classification datasets and robustness to label corruption.

Evidence boundary. Edit-path interpolation and learned diffusion corruption are analogous trajectory designs, not the same stochastic process.

Chronology. 2023-06-02: arxiv version.

Candidate dates. 2023-06-02: arXiv 2306.01310v1; 2024-04-16: IJCAI conference notification.

Domains. General ML (central); Molecules & drug discovery (evaluation); Proteins & genomics (evaluation).

Methodologies. Graph neural networks (used); Data augmentation (used).

Concepts. Graph structure & transformations (central); Adaptive trajectories (conceptual parallel); Constraint-aware design (used).

Keywords. Adaptive trajectories; Constraint-aware design; context-sensitive cost; Data augmentation; graph edit path; Graph neural networks; Graph structure & transformations; Graphs & discrete problems; interpolation; label corruption.

Sources. arXiv 2306.01310v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Difficult, not too different — Learned, constrained augmentation (documented); see r017.
• MELD — Choose informative transformation paths (interpretive); see r019.

LED-GFN

Learning Energy Decompositions for Partial Inference in GFlowNets

ICLR 2024 · Review: fulltext.

Provides local training signals when intermediate object energies are costly or misleading.

Contribution. Learned, smoothly changing potential functions decompose terminal energy while preserving the target optimal policy.

Map contribution. Search & optimization: Provides local energy signals for reward-guided search while preserving the target sampling distribution.

Evaluation. Bags, molecular graphs, RNA, sets, and maximum independent sets.

Evidence boundary. The guarantee concerns the specified flow reparameterization; learned local potentials need not be physical energies.

Other contributions. Generative modeling.

Chronology. 2023-10-05: arxiv version.

Candidate dates. 2023-10-05: arXiv 2310.03301v1; 2024-01-15: ICLR conference notification.

Domains. General ML (central); Molecules & drug discovery (central); Proteins & genomics (central).

Methodologies. GFlowNets (used); Probabilistic inference (used).

Concepts. Credit assignment (central); Factorization & decomposition (central); Local feedback (central); Amortized inference / generation (used).

Keywords. Amortized inference / generation; Credit assignment; Factorization & decomposition; General learning & inference; GFlowNets; Graphs & discrete problems; Local feedback; maximum independent set; Molecules & drug discovery; potential function; Probabilistic inference; Proteins & genomics; RNA design; terminal energy decomposition.

Sources. arXiv 2310.03301v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Pessimistic backward policy — Improve GFlowNet training without changing its target (documented); see r049.
• Co-folding representations — Dense feedback for molecular optimization (interpretive); see r052.
• Latent veracity inference — Locate feedback inside a long construction (interpretive); see r053.

Gap-encoded edge lists

A Simple and Scalable Representation for Graph Generation

ICLR 2024 · Review: fulltext.

Makes sparse graph generation scale with edges rather than full adjacency matrices.

Contribution. GEEL combines edge lists, gap encoding, bandwidth restrictions, and an attributed-graph grammar.

Map contribution. Generative modeling: A compact edge grammar makes sparse graph generation scale with the edge count.

Evaluation. Ten non-attributed and two molecular graph-generation tasks.

Evidence boundary. Compactness depends on graph structure and ordering; the representation is not a universal constant-size graph description.

Other contributions. Representations.

Chronology. 2023-12-04: arxiv version.

Candidate dates. 2023-12-04: arXiv 2312.02230v1; 2024-01-15: ICLR conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. Language / sequence models (used); Compression & compact coding (used).

Concepts. Structured tokenization (central); Compression (central); Graph structure & transformations (central).

Keywords. attributed grammar; bandwidth; Compression; Compression & compact coding; edge list; gap encoding; GEEL; Graph structure & transformations; Graphs & discrete problems; Language / sequence models; Molecules & drug discovery; Structured tokenization.

Sources. arXiv 2312.02230v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Spanning-tree molecular generation — Compact graph construction grammars (documented); see r041.
• K²-tree graph generation — Sparse graph tokenization (documented); see r043.
• DNAChunker — Choose tokens that respect the object (interpretive); see r044.

K²-tree graph generation

Graph Generation with K^2 Trees

ICLR 2024 · Review: fulltext.

Generates graphs through a compact hierarchical representation of their adjacency structure.

Contribution. A tokenized K²-tree and tree-aware Transformer exploit sparse submatrices for sequential generation.

Map contribution. Generative modeling: A hierarchical adjacency grammar supports compact graph generation.

Evaluation. Four general graph and two molecular graph datasets.

Evidence boundary. A useful compression grammar differs from a learned molecular representation or a chemical reaction pathway.

Other contributions. Representations.

Chronology. 2023-05-30: arxiv version.

Candidate dates. 2023-05-30: arXiv 2305.19125v1; 2024-01-15: ICLR conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. Language / sequence models (used); Compression & compact coding (used).

Concepts. Structured tokenization (central); Compression (central); Adaptive resolution (central); Graph structure & transformations (central).

Keywords. Adaptive resolution; Compression; Compression & compact coding; Graph structure & transformations; Graphs & discrete problems; hierarchical tokens; K²-tree; Language / sequence models; Molecules & drug discovery; sparse adjacency; Structured tokenization; tree-aware Transformer.

Sources. arXiv 2305.19125v4: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Spanning-tree molecular generation — Hierarchical structure for graph generation (documented); see r042.
• Gap-encoded edge lists — Sparse graph tokenization (documented); see r043.

Local Search GFlowNets

Local Search GFlowNets

ICLR 2024 · Review: fulltext.

Combines global sampling with refinement of promising local neighborhoods.

Contribution. Backward-policy backtracking and forward reconstruction produce improved trajectories for replay training.

Map contribution. Search & optimization: Refines promising objects through trajectory backtracking, reconstruction, and replay.

Evaluation. Molecule optimization and biological-sequence design, measuring reward-distribution accuracy and discovered modes.

Evidence boundary. GFlowNets target reward-proportional sampling, whereas genetic expert learning principally pursues high rewards; their objectives are not interchangeable.

Other contributions. Generative modeling.

Chronology. 2023-10-04: arxiv version.

Candidate dates. 2023-10-04: arXiv 2310.02710v1; 2024-01-15: ICLR conference notification.

Domains. General ML (central); Molecules & drug discovery (central); Proteins & genomics (central).

Methodologies. GFlowNets (used); Search & optimization (used).

Concepts. Exploration–exploitation (central); Local search (central); Diversity & mode coverage (central); Replay & off-policy reuse (central); Search–learning feedback (central).

Keywords. backtracking; Diversity & mode coverage; Exploration–exploitation; forward reconstruction; General learning & inference; GFlowNets; Local search; Molecules & drug discovery; Proteins & genomics; Replay & off-policy reuse; reward-proportional sampling; Search & optimization; Search–learning feedback; trajectory refinement.

Sources. arXiv 2310.02710v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Genetic expert-guided learning — Exploration followed by local refinement (interpretive); see r031.
• RxnFlow — Reward-proportional molecular search (documented); see r039.
• Pessimistic backward policy — Backward trajectories affect discovery (documented); see r050.
• Adaptive Teachers — Coverage through additional training trajectories (documented); see r051.

Wavelet graph diffusion

Multi-resolution Spectral Coherence for Graph Generation with Score-based Diffusion

NeurIPS 2023 · Review: fulltext.

Preserves graph-frequency structure during joint node-and-edge generation.

Contribution. Multi-resolution graph wavelets couple node and edge signals within score-based diffusion.

Map contribution. Generative modeling: Preserves graph-frequency structure during joint node-and-edge generation.

Evaluation. Four representative graph-generation benchmarks and multi-resolution analyses.

Evidence boundary. A spectral graph signal is not a physical particle-density field; the connection is representational.

Other contributions. Representations.

Chronology. 2023-09-22: conference notification.

Candidate dates. 2023-09-22: NeurIPS conference notification.

Domains. General ML (central); Molecules & drug discovery (evaluation).

Methodologies. Diffusion models (used); Graph neural networks (used).

Concepts. Adaptive resolution (central); Graph structure & transformations (central); Density / field representations (conceptual parallel).

Keywords. Adaptive resolution; Density / field representations; Diffusion models; Graph neural networks; Graph structure & transformations; graph wavelet; Graphs & discrete problems; node–edge coupling; score-based generation; spectral signal.

Sources. Conference proceedings: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Structured node diffusion — Graph diffusion beyond independent variables (documented); see r045.
• Gaussian plane-wave neural operator — Separate coarse and fine spatial information (interpretive); see r046.

Structured node diffusion

Diffusion Probabilistic Models for Structured Node Classification

NeurIPS 2023 · Review: fulltext.

Predicts graph-node labels jointly rather than independently.

Contribution. A partially supervised diffusion objective and constrained sampling incorporate observed labels into the remaining predictions.

Map contribution. Generative modeling: Conditional diffusion models coupled node-label predictions.

Evaluation. Partially labeled, unlabeled, transductive, and inductive node-classification settings.

Evidence boundary. The expressivity analysis is specific to the proposed structured prediction construction and AGG-WL comparison.

Other contributions. Learning & inference.

Chronology. 2023-02-21: arxiv version.

Candidate dates. 2023-02-21: arXiv 2302.10506v1; 2023-09-22: NeurIPS conference notification.

Domains. General ML (central); Proteins & genomics (evaluation).

Methodologies. Diffusion models (used); Graph neural networks (used); Probabilistic inference (used).

Concepts. Conditional generation (central); Message passing (central); Constraint-aware design (used).

Keywords. AGG-WL; Conditional generation; constrained sampling; Constraint-aware design; Diffusion models; Graph neural networks; Graphs & discrete problems; Message passing; partially observed labels; Probabilistic inference; structured prediction.

Sources. arXiv 2302.10506v5: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Non-backtracking GNNs — Beyond independent local node decisions (documented); see r021.
• Wavelet graph diffusion — Graph diffusion beyond independent variables (documented); see r045.

BootGen

Bootstrapped Training of Score-Conditioned Generator for Offline Design of Biological Sequences

NeurIPS 2023 · Review: fulltext.

Designs biological sequences from a fixed offline dataset of scored examples.

Contribution. Rank-weighted training alternates with proxy-labeled generated data, then aggregates models for diversity.

Map contribution. Search & optimization: Proxy-guided bootstrapping searches for improved designs from a fixed dataset.

Evaluation. Six RNA, DNA, and protein design tasks with component ablations.

Evidence boundary. Proxy-labeled bootstrapping can inherit proxy errors; offline optimization does not supply new ground-truth assays during search.

Other contributions. Generative modeling.

Chronology. 2023-06-05: arxiv version.

Candidate dates. 2023-06-05: arXiv 2306.03111v1; 2023-09-22: NeurIPS conference notification.

Domains. Proteins & genomics (central).

Methodologies. Representation learning (used); Search & optimization (used).

Concepts. Self-training (central); Diversity & mode coverage (central); Simulation / proxy-to-reality gap (central); Knowledge distillation (central); Amortized inference / generation (used).

Keywords. Amortized inference / generation; Diversity & mode coverage; Knowledge distillation; model aggregation; offline design; Proteins & genomics; proxy-labeled data; rank weighting; Representation learning; Search & optimization; Self-training; Simulation / proxy-to-reality gap.

Sources. arXiv 2306.03111v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• RoMA — Offline design with a learned proxy (documented); see r061.
• CORE-PO — Pseudo-labels shape the next learner (interpretive); see r063.

Concept intervention analysis

A Closer Look at the Intervention Procedure of Concept Bottleneck Models

ICML 2023 · Review: fulltext.

Examines how concept bottleneck interventions should be selected and applied.

Contribution. Controlled studies vary intervention strategy, granularity, and environment to expose reliability and fairness pitfalls.

Map contribution. Benchmarks: Controlled evaluation identifies reliability and fairness limits of concept interventions.

Evaluation. CUB, SkinCon, and synthetic causal datasets under matched intervention budgets.

Evidence boundary. A concept edit is not automatically a valid causal intervention; effectiveness depends on the intervention protocol.

Chronology. 2023-02-28: arxiv version.

Candidate dates. 2023-02-28: arXiv 2302.14260v1; 2023-04-24: ICML conference notification.

Domains. General ML (central).

Methodologies. Representation learning (used); Benchmarks & evaluation (used).

Concepts. Verification & diagnostics (central); Causal reasoning & interventions (central); Local feedback (central); Benchmarking (central).

Keywords. Benchmarking; Benchmarks & evaluation; Causal reasoning & interventions; concept bottleneck; CUB; fairness; General learning & inference; intervention budget; Local feedback; Representation learning; SkinCon; Verification & diagnostics.

Sources. arXiv 2302.14260v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Causal Influence Prompting — An interpretable edit needs a valid intervention protocol (interpretive); see r025.

Planner-guided imitation

Imitating Graph-Based Planning with Goal-Conditioned Policies

ICLR 2023 · Review: fulltext.

Transfers graph-planner knowledge into a goal-conditioned control policy.

Contribution. Subgoal-conditioned policies teach target-goal policies, while stochastic subgoal skipping improves execution.

Map contribution. Search & optimization: Policies learn reusable control behavior from subgoal-planner results.

Evaluation. Control performance, planner-free execution, imitation alternatives, and subgoal-skipping ablations.

Evidence boundary. Planner distillation is demonstrated in long-horizon control tasks rather than general real-world robotics transfer.

Other contributions. Learning & inference.

Chronology. 2023-01-21: conference notification.

Candidate dates. 2023-01-21: ICLR conference notification; 2023-03-20: arXiv 2303.11166v1.

Domains. General ML (central).

Methodologies. Reinforcement learning (used); Search & optimization (used); Representation learning (used).

Concepts. Knowledge distillation (central); Search–learning feedback (central); Sample efficiency (central).

Keywords. Control & multi-agent learning; goal-conditioned policy; graph planner; Knowledge distillation; planner-free execution; Reinforcement learning; Representation learning; Sample efficiency; Search & optimization; Search–learning feedback; subgoal skipping.

Sources. arXiv 2303.11166v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Variational information distillation — Distilling an expensive teacher (documented); see r009.
• Self-improved retrosynthesis — Amortizing planner knowledge (documented); see r033.
• Adaptive-grid exploration — Subgoal structure for long-horizon control (documented); see r034.

Learning with a biased committee

Learning Debiased Classifier with Biased Committee

NeurIPS 2022 · Review: fulltext.

Finds bias-conflicting training examples without spurious-attribute labels.

Contribution. A diverse biased committee identifies hard examples and exchanges knowledge with the main classifier as training progresses.

Map contribution. Learning & inference: Finds bias-conflicting training examples without spurious-attribute labels.

Evaluation. Five real-world debiasing datasets, with comparisons using and omitting bias labels.

Evidence boundary. Self-supervised pretraining increases cost and ensemble randomness can affect performance.

Chronology. 2022-06-22: arxiv version.

Candidate dates. 2022-06-22: arXiv 2206.10843v1; 2022-09-14: NeurIPS conference notification.

Domains. General ML (central).

Methodologies. Robust learning (used); Representation learning (used).

Concepts. Auxiliary teachers (central); Signal–noise separation (central); Curriculum & training difficulty (central); Invariance & spurious factors (central).

Keywords. Auxiliary teachers; bias-conflicting samples; biased committee; Curriculum & training difficulty; General learning & inference; Invariance & spurious factors; knowledge exchange; Representation learning; Robust learning; self-supervised pretraining; Signal–noise separation.

Sources. arXiv 2206.10843v5: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Learning from Failure — Bias models expose difficult examples (documented); see r013.
• Multi-bias robust learning — Spurious-correlation mitigation (documented); see r014.

Disentangled risk-sensitive MARL

Disentangling Sources of Risk for Distributional Multi-Agent Reinforcement Learning

ICML 2022 · Review: fulltext.

Separates uncertainty from teammates’ actions and uncertainty from environmental dynamics.

Contribution. True and transformed distributional value estimators allow distinct risk sensitivities for environment and cooperation.

Map contribution. Learning & inference: Separate value distributions model uncertainty from teammates and the environment.

Evaluation. Controlled multi-agent games and cooperative benchmark scenarios with risk-sensitivity studies.

Evidence boundary. Multi-agent RL policies are different objects from debating language-model agents; shared multi-agent terminology does not imply the same algorithm.

Chronology. 2022-05-14: conference notification.

Candidate dates. 2022-05-14: ICML conference notification.

Domains. General ML (central).

Methodologies. Reinforcement learning (used); Robust learning (used).

Concepts. Uncertainty & risk (central); Factorization & decomposition (central); Exploration–exploitation (central).

Keywords. Control & multi-agent learning; distributional value; environment uncertainty; Exploration–exploitation; Factorization & decomposition; Reinforcement learning; risk sensitivity; Robust learning; teammate uncertainty; Uncertainty & risk.

Sources. Conference proceedings: Abstract; method formulation; experimental results and limitations.

Specific connections.

• INDIBATOR — Multiple agents contribute different uncertainty or viewpoints (interpretive); see r036.

Difficult, not too different

What Makes Better Augmentation Strategies? Augment Difficult but Not Too Different

ICLR 2022 · Review: partial.

Learns text augmentations that are challenging while retaining original semantics.

Contribution. An augmentation policy balances low model confidence and high semantic similarity, with sample reweighting.

Map contribution. Learning & inference: Learns text augmentations that are challenging while retaining original semantics.

Evaluation. Accessible sources describe text classification, GLUE, and low-data or imbalanced settings.

Evidence boundary. Only indexed publisher excerpts and the author implementation were accessible; fine experimental comparisons remain partially reviewed.

No arXiv link is listed for this work. This draft uses the accessible abstract and, where available, author implementation; full-paper review remains pending.

Chronology. 2022-01-24: conference notification.

Candidate dates. 2022-01-24: ICLR conference notification.

Domains. General ML (central).

Methodologies. Data augmentation (used); Reinforcement learning (used); Representation learning (used).

Concepts. Curriculum & training difficulty (central); Constraint-aware design (central); Signal–noise separation (central).

Keywords. Constraint-aware design; Curriculum & training difficulty; Data augmentation; GLUE; Language & scientific reasoning; learned augmentation; low confidence; Reinforcement learning; Representation learning; sample reweighting; semantic similarity; Signal–noise separation.

Sources. Author implementation: README and indexed abstract; partial review.

Specific connections.

• EPIC — Learned, constrained augmentation (documented); see r017.
• Adaptive Teachers — Informative difficulty as a curriculum (interpretive); see r018.

Spanning-tree molecular generation

Spanning Tree-based Graph Generation for Molecules

ICLR 2022 · Review: partial.

Generates molecular graphs by constructing a spanning tree and its remaining edges.

Contribution. Compact tree operations exploit sparsity and permit structural constraints such as valence rules.

Map contribution. Generative modeling: Generates molecular graphs by constructing a spanning tree and its remaining edges.

Evaluation. The conference abstract reports QM9, ZINC250K, MOSES, and penalized-logP optimization; fine experimental comparisons remain partially reviewed.

Evidence boundary. Only indexed publisher abstract and introduction excerpts were accessible; full experimental and limitation review remains pending.

No arXiv link is listed for this work. This draft uses the accessible abstract and, where available, author implementation; full-paper review remains pending.

Chronology. 2022-01-24: conference notification.

Candidate dates. 2022-01-24: ICLR conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. Language / sequence models (used); Search & optimization (used).

Concepts. Structured tokenization (central); Constraint-aware design (central); Graph structure & transformations (central).

Keywords. compact tree operation; Constraint-aware design; Graph structure & transformations; Graphs & discrete problems; Language / sequence models; Molecules & drug discovery; residual edge; Search & optimization; spanning tree; Structured tokenization; tree-relative position; valence.

Sources. ICLR conference abstract: Conference abstract; partial review.

Specific connections.

• RxnFlow — Feasibility built into the action space (documented); see r040.
• Gap-encoded edge lists — Compact graph construction grammars (documented); see r041.
• K²-tree graph generation — Hierarchical structure for graph generation (documented); see r042.

RoMA

RoMA: Robust Model Adaptation for Offline Model-based Optimization

NeurIPS 2021 · Review: fulltext.

Optimizes a learned objective while limiting exploitation of out-of-distribution model errors.

Contribution. Robust proxy pretraining and candidate-specific adaptation use a local smoothness prior during gradient-based design.

Map contribution. Search & optimization: Optimizes a learned objective while limiting exploitation of out-of-distribution model errors.

Evaluation. Six Design-bench offline tasks, including biological and control-design applications.

Evidence boundary. Local smoothness is a modeling assumption, and avoiding proxy overestimation does not guarantee every optimized design is valid.

Other contributions. Learning & inference.

Chronology. 2021-09-28: conference notification.

Candidate dates. 2021-09-28: NeurIPS conference notification; 2021-10-27: arXiv 2110.14188v1.

Domains. General ML (central); Proteins & genomics (central).

Methodologies. Robust learning (used); Search & optimization (used).

Concepts. Simulation / proxy-to-reality gap (central); Local search (central); Constraint-aware design (central).

Keywords. Constraint-aware design; Control & multi-agent learning; Design-bench; General learning & inference; Local search; local smoothness; offline model-based optimization; Proteins & genomics; proxy overestimation; Robust learning; Search & optimization; Simulation / proxy-to-reality gap.

Sources. arXiv 2110.14188v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• BootGen — Offline design with a learned proxy (documented); see r061.
• Structurally diverse molecular LLMs — A proxy improvement can miss the intended goal (interpretive); see r062.

Self-improved retrosynthesis

Self-Improved Retrosynthetic Planning

ICML 2021 · Review: fulltext.

Trains retrosynthetic reaction proposals to support complete executable search routes.

Contribution. The reaction model imitates successful routes found by its planner, with forward-model reaction augmentation.

Map contribution. Search & optimization: Successful planner routes train reaction proposals for later synthesis searches.

Evaluation. Retrosynthetic planning success and reaction prediction, with self-improvement and augmentation studies.

Evidence boundary. Routes use available building blocks and modeled reactions; computational success does not independently establish laboratory yield.

Other contributions. Generative modeling.

Chronology. 2021-05-08: conference notification.

Candidate dates. 2021-05-08: ICML conference notification; 2021-06-09: arXiv 2106.04880v1.

Domains. Molecules & drug discovery (central).

Methodologies. Search & optimization (used); Representation learning (used); Data augmentation (used).

Concepts. Synthesis & reactant availability (central); Search–learning feedback (central); Self-training (central); Constraint-aware design (central).

Keywords. available reactants; Constraint-aware design; Data augmentation; forward reaction augmentation; Molecules & drug discovery; Representation learning; retrosynthesis; Search & optimization; Search–learning feedback; Self-training; successful route imitation; Synthesis & reactant availability.

Sources. arXiv 2106.04880v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Genetic expert-guided learning — Search teaches the next model (documented); see r032.
• Planner-guided imitation — Amortizing planner knowledge (documented); see r033.
• RetCL — Retrosynthesis under available reactants (documented); see r037.
• Latent veracity inference — Successful search supplies reusable supervision (interpretive); see r076.

RetCL

RetCL: A Selection-based Approach for Retrosynthesis via Contrastive Learning

IJCAI 2021 · Review: fulltext.

Selects available reactants rather than generating unrestricted reactant descriptions.

Contribution. Graph embeddings and contrastive hard-negative training rank combinations from a candidate inventory.

Map contribution. Representations: Contrastive graph embeddings rank available reactant combinations.

Evaluation. USPTO retrosynthesis, large candidate sets, and generalization to unseen reaction templates.

Evidence boundary. Availability is defined by the chosen candidate library; commercial selection does not prove a reaction succeeds.

Other contributions. Search & optimization.

Chronology. 2021-04-30: conference notification.

Candidate dates. 2021-04-30: IJCAI conference notification; 2021-05-03: arXiv 2105.00795v1.

Domains. Molecules & drug discovery (central).

Methodologies. Representation learning (used); Graph neural networks (used); Search & optimization (used).

Concepts. Synthesis & reactant availability (central); Constraint-aware design (central); Contrastive learning (central).

Keywords. candidate inventory; Constraint-aware design; Contrastive learning; Graph neural networks; hard negatives; Molecules & drug discovery; reactant selection; Representation learning; Search & optimization; Synthesis & reactant availability; unseen reaction template; USPTO.

Sources. arXiv 2105.00795v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Self-improved retrosynthesis — Retrosynthesis under available reactants (documented); see r037.
• RxnFlow — Availability constrains molecular design (documented); see r038.

Layer-adaptive pruning

Layer-adaptive sparsity for the Magnitude-based Pruning

ICLR 2021 · Review: fulltext.

Allocates sparsity across neural-network layers without extensive layerwise tuning.

Contribution. A rescaled magnitude score accounts for model-level distortion when choosing which weights to prune.

Map contribution. Learning & inference: Layer-specific pruning scores control model distortion without extensive tuning.

Evaluation. Image classification, one-shot pruning, weight rewinding, and layerwise-sparsity comparisons.

Evidence boundary. The distortion argument and empirical pruning findings concern the tested architectures and schedules.

Chronology. 2020-10-15: arxiv version.

Candidate dates. 2020-10-15: arXiv 2010.07611v1; 2021-01-14: ICLR conference notification.

Domains. General ML (central).

Methodologies. Compression & compact coding (used).

Concepts. Compression (central); Adaptive resolution (conceptual parallel).

Keywords. Adaptive resolution; Compression; Compression & compact coding; General learning & inference; layerwise sparsity; magnitude pruning; model distortion; weight rewinding.

Sources. arXiv 2010.07611v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• DNAChunker — Adaptive allocation of representational budget (interpretive); see r012.

Learning from Failure

Learning from Failure: Training Debiased Classifier from Biased Classifier

NeurIPS 2020 · Review: fulltext.

Debiases a classifier by emphasizing examples that an intentionally biased model gets wrong.

Contribution. Two coupled learners exploit differences in learning difficulty between spurious and intended attributes.

Map contribution. Learning & inference: Debiases a classifier by emphasizing examples that an intentionally biased model gets wrong.

Evaluation. Synthetic biased data and real-world debiasing tasks without explicit bias-attribute labels.

Evidence boundary. The useful difficulty cue assumes spurious correlations are comparatively easy to learn in the studied setting.

Chronology. 2020-07-06: arxiv version.

Candidate dates. 2020-07-06: arXiv 2007.02561v1; 2020-09-26: NeurIPS conference notification.

Domains. General ML (central).

Methodologies. Robust learning (used).

Concepts. Auxiliary teachers (central); Signal–noise separation (central); Curriculum & training difficulty (central); Invariance & spurious factors (central).

Keywords. Auxiliary teachers; biased learner; Curriculum & training difficulty; failure weighting; General learning & inference; Invariance & spurious factors; learning difficulty; Robust learning; Signal–noise separation; spurious correlation.

Sources. arXiv 2007.02561v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Learning with a biased committee — Bias models expose difficult examples (documented); see r013.
• Adaptive Teachers — An auxiliary model redirects training (interpretive); see r015.

Genetic expert-guided learning

Guiding Deep Molecular Optimization with Genetic Exploration

NeurIPS 2020 · Review: fulltext.

Uses domain-specific genetic search to improve a neural molecular generator.

Contribution. Mutation and crossover refine generated molecules, then an apprentice imitates high-reward samples from priority queues.

Map contribution. Search & optimization: Uses domain-specific genetic search to improve a neural molecular generator.

Evaluation. Penalized logP, similarity-constrained optimization, GuacaMol, and filtered molecule analyses.

Evidence boundary. High scores on molecular proxies do not establish viable drugs; the paper also examines post-hoc chemical filtering.

Other contributions. Generative modeling.

Chronology. 2020-07-04: arxiv version.

Candidate dates. 2020-07-04: arXiv 2007.04897v1; 2020-09-26: NeurIPS conference notification.

Domains. Molecules & drug discovery (central).

Methodologies. Search & optimization (used); Representation learning (used).

Concepts. Genetic algorithms (central); Search–learning feedback (central); Local search (central); Exploration–exploitation (central); Replay & off-policy reuse (central); Simulation / proxy-to-reality gap (evaluation).

Keywords. apprentice; crossover; Exploration–exploitation; GEGL; Genetic algorithms; Local search; Molecules & drug discovery; mutation; penalized logP; priority queue; Replay & off-policy reuse; Representation learning; Search & optimization; Search–learning feedback; Simulation / proxy-to-reality gap.

Sources. arXiv 2007.04897v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Symmetric replay training — Reuse good candidates under expensive rewards (documented); see r030.
• Local Search GFlowNets — Exploration followed by local refinement (interpretive); see r031.
• Self-improved retrosynthesis — Search teaches the next model (documented); see r032.
• MaskGXT / HACO — A specialized search partner improves a generator (interpretive); see r088.

Learning what to defer

Learning What to Defer for Maximum Independent Sets

ICML 2020 · Review: fulltext.

Solves large maximum-independent-set problems with an adaptive number of decision stages.

Contribution. An agent selects, rejects, or defers vertex decisions and can combine its solution with local search.

Map contribution. Search & optimization: Solves large maximum-independent-set problems with an adaptive number of decision stages.

Evaluation. Synthetic and real-world large graphs and additional locally decomposable combinatorial tasks.

Evidence boundary. Performance comparisons depend on graph families and time budgets; the learned solver does not certify optimality for arbitrary MIS instances.

Chronology. 2020-06-01: conference notification.

Candidate dates. 2020-06-01: ICML conference notification; 2020-06-17: arXiv 2006.09607v1.

Domains. General ML (central).

Methodologies. Reinforcement learning (used); Graph neural networks (used); Search & optimization (used).

Concepts. Combinatorial optimization (central); Constraint-aware design (central); Adaptive trajectories (central); Sample efficiency (central); Local search (used).

Keywords. adaptive stages; Adaptive trajectories; Combinatorial optimization; Constraint-aware design; deferred decisions; Graph neural networks; Graphs & discrete problems; Local search; locally decomposable objective; maximum independent set; Reinforcement learning; Sample efficiency; Search & optimization.

Sources. arXiv 2006.09607v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• RL4CO — Learning discrete solvers and judging them fairly (documented); see r026.
• Unsupervised combinatorial optimization — Constrained discrete decisions (documented); see r027.

Variational information distillation

Variational Information Distillation for Knowledge Transfer

CVPR 2019 · Review: fulltext.

Transfers knowledge between neural networks by maximizing shared information.

Contribution. A variational mutual-information objective supports distillation across heterogeneous architectures.

Map contribution. Representations: A variational information objective transfers representations between different neural architectures.

Evaluation. Image-classification distillation, transfer learning, and CNN-to-MLP knowledge transfer.

Evidence boundary. The practical information estimator uses a particular observation model; transferred features need not retain every task-relevant property.

Chronology. 2019-03-02: conference notification.

Candidate dates. 2019-03-02: CVPR conference notification; 2019-04-11: arXiv 1904.05835v1.

Domains. General ML (central).

Methodologies. Representation learning (used); Probabilistic inference (used); Compression & compact coding (used).

Concepts. Knowledge distillation (central); Information routing (used); Compression (central).

Keywords. CNN-to-MLP; Compression; Compression & compact coding; General learning & inference; heterogeneous architecture; Information routing; Knowledge distillation; mutual information; Probabilistic inference; Representation learning; variational bound.

Sources. arXiv 1904.05835v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Bucket renormalization — Compressing useful information (interpretive); see r008.
• Planner-guided imitation — Distilling an expensive teacher (documented); see r009.
• Co-folding representations — Representation transfer (documented); see r010.
• scTrilemma — What information should an embedding retain? (interpretive); see r011.

Bucket renormalization

Bucket-Renormalization for Approximate Inference

ICML 2018 · JSTAT 2019 · Review: fulltext.

Approximates graphical-model partition functions through sequential variable elimination.

Contribution. Low-rank mini-bucket projections borrow tensor-network renormalization ideas and can be globally calibrated.

Map contribution. Learning & inference: Low-rank factor projections control the cost of partition-function inference.

Evaluation. Ising models and UAI inference benchmarks, comparing mini-bucket and variational methods.

Evidence boundary. Approximation quality depends on decomposition and rank choices; convergence-free elimination does not mean exact general inference.

Other contributions. Representations.

Chronology. 2018-03-14: arxiv version.

Candidate dates. 2018-03-14: arXiv 1803.05104v1; 2018-05-12: ICML conference notification.

Domains. General ML: graphical models (central).

Methodologies. Probabilistic inference (used).

Concepts. Factorization & decomposition (central); Compression (central); Unnormalized distributions (central); Learned / classical hybrids (conceptual parallel).

Keywords. Compression; Factorization & decomposition; General learning & inference; Learned / classical hybrids; low-rank projection; mini-bucket; partition function; Probabilistic inference; tensor network; Unnormalized distributions; variable elimination.

Sources. arXiv 1803.05104v3: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Gauged mini-bucket elimination — Bounded elimination and factor compression (documented); see r004.
• Variational information distillation — Compressing useful information (interpretive); see r008.

Gauged variational inference

Gauging Variational Inference

NeurIPS 2017 · JSTAT 2019 · Review: fulltext.

Improves partition-function approximations by transforming graphical-model factors.

Contribution. Optimized gauge transformations preserve the partition function while improving mean-field and belief-propagation bounds.

Map contribution. Learning & inference: Optimized gauge transformations improve mean-field and belief-propagation bounds for partition-function inference.

Evaluation. Small complete models and larger graphical models with controlled interaction strengths.

Evidence boundary. Exactness results require special graph structure; general models remain approximate and optimization can be costly.

Chronology. 2017-03-03: arxiv version.

Candidate dates. 2017-03-03: arXiv 1703.01056v1; 2017-09-04: NeurIPS conference notification.

Domains. General ML: graphical models (central).

Methodologies. Probabilistic inference (used).

Concepts. Unnormalized distributions (central); Symmetry (central); Factorization & decomposition (central).

Keywords. Factorization & decomposition; Forney graph; gauge transformation; General learning & inference; mean field; partition function; Probabilistic inference; Symmetry; Unnormalized distributions.

Sources. arXiv 1703.01056v5: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Gauged mini-bucket elimination — Gauge-preserving inference (documented); see r003.
• MCMC + belief propagation — Improving partition estimates (documented); see r005.

Gauged mini-bucket elimination

Gauged Mini-Bucket Elimination for Approximate Inference

AISTATS 2018 · Review: fulltext.

Combines gauge transformations with bounded-complexity variable elimination.

Contribution. Joint optimization tightens weighted mini-bucket upper or lower bounds, with improved optimization behavior.

Map contribution. Learning & inference: Joint gauge and weighted mini-bucket optimization improves upper and lower partition-function bounds.

Evaluation. Grid Ising, Forney-style graphical models, and UAI linkage instances.

Evidence boundary. Bound direction and guarantees depend on the stated weighted-elimination formulation and model assumptions.

Chronology. 2017-12-22: conference notification.

Candidate dates. 2017-12-22: AISTATS conference notification; 2018-01-05: arXiv 1801.01649v1.

Domains. General ML: graphical models (central).

Methodologies. Probabilistic inference (used).

Concepts. Unnormalized distributions (central); Symmetry (central); Factorization & decomposition (central).

Keywords. Factorization & decomposition; gauge optimization; General learning & inference; lower bound; Probabilistic inference; Symmetry; Unnormalized distributions; upper bound; weighted mini-bucket.

Sources. arXiv 1801.01649v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Gauged variational inference — Gauge-preserving inference (documented); see r003.
• Bucket renormalization — Bounded elimination and factor compression (documented); see r004.

Odd-cycle matching BP

Maximum Weight Matching using Odd-sized Cycles: Max-Product Belief Propagation and Half-Integrality

IEEE TIT 2018 · Review: fulltext.

Reformulates maximum-weight matching so belief propagation can exploit odd-cycle constraints.

Contribution. Graph transformations restore convergence under tight LP relaxations; half-integral solutions support iterative cutting-plane heuristics.

Map contribution. Search & optimization: Graph transformations and cutting-plane steps solve constrained matching problems.

Evaluation. Matching instances and comparisons of BP-based and LP-based cutting-plane procedures.

Evidence boundary. The guarantee requires the stated tightness and uniqueness conditions; the arXiv manuscript has an earlier title than the journal record.

Other contributions. Learning & inference.

Chronology. 2018-01-01: arxiv version. The map uses the first arXiv version listing Sungsoo Ahn as a coauthor. Earlier versions without him are excluded.

Candidate dates. 2018-01-01: arXiv 1306.1167v2; 2018-03: IEEE TIT journal month.

Excluded versions. 1306.1167v1 (2013-06-05): Sungsoo Ahn is absent from this version’s author list.

Domains. General ML: graphical models (central).

Methodologies. Probabilistic inference (used); Search & optimization (used).

Concepts. Combinatorial optimization (central); Message passing (central); Graph structure & transformations (central); Constraint-aware design (central).

Keywords. Combinatorial optimization; Constraint-aware design; cutting plane; General learning & inference; Graph structure & transformations; Graphs & discrete problems; half-integrality; LP relaxation; maximum-weight matching; Message passing; odd-cycle constraint; Probabilistic inference; Search & optimization.

Sources. arXiv 1306.1167v2: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Blossom belief propagation — Matching through graph transformations (documented); see r001.
• MCMC + belief propagation — Repairing an approximate inference engine (interpretive); see r002.

MCMC + belief propagation

Synthesis of MCMC and Belief Propagation

NeurIPS 2016 · Review: fulltext.

Corrects belief-propagation partition estimates using Monte Carlo on a transformed loop model.

Contribution. BP-aware Markov chains sample loop-calculus corrections, combining a fast approximation with stochastic error correction.

Map contribution. Learning & inference: Monte Carlo loop corrections improve belief-propagation partition estimates.

Evaluation. Ising and hard-core models with BP, Gibbs sampling, and loop-based MCMC comparisons.

Evidence boundary. Truncated and full loop-series schemes have different guarantees; faster mixing is not asserted for every graphical model.

Chronology. 2016-05-29: arxiv version. The official reviewer instructions place notification in mid-September 2016, after the May arXiv posting; the exact notification day is not archived.

Candidate dates. 2016-05-29: arXiv 1605.09042v1.

Domains. General ML: graphical models (central).

Methodologies. Probabilistic inference (used); Energy-based sampling (used).

Concepts. MCMC (central); Message passing (central); Unnormalized distributions (central); Learned / classical hybrids (central).

Keywords. BP correction; Energy-based sampling; General learning & inference; hard-core model; Ising; Learned / classical hybrids; loop calculus; MCMC; Message passing; partition function; Probabilistic inference; Unnormalized distributions.

Sources. arXiv 1605.09042v6: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Odd-cycle matching BP — Repairing an approximate inference engine (interpretive); see r002.
• Gauged variational inference — Improving partition estimates (documented); see r005.
• Iterated energy-based flow matching — Unnormalized targets, different estimators (interpretive); see r006.
• Search-guided diffusion samplers — Approximation with a stochastic search partner (interpretive); see r007.
• MADField — A fast approximation benefits from another solver (interpretive); see r082.

Blossom belief propagation

Minimum Weight Perfect Matching via Blossom Belief Propagation

NeurIPS 2015 · Review: fulltext.

Solves minimum-weight perfect matching through successive belief-propagation problems.

Contribution. Blossom contraction and expansion make BP a distributed subroutine in a polynomial-run matching algorithm.

Map contribution. Search & optimization: Blossom contraction and expansion turn belief propagation into a polynomial-time matching solver.

Evaluation. Theoretical termination, correctness, and interpretation through successive linear programs.

Evidence boundary. The result is specific to matching; it does not establish polynomial algorithms for arbitrary graphical-model MAP problems.

Other contributions. Learning & inference.

Chronology. 2015-09-23: arxiv version. The archived NIPS 2015 calendar omits the paper notification date. Its CFP gives a review-period end of 2015-09-04, not a separately verified acceptance announcement. The verified arXiv date is used pending confirmation; this does not change its order relative to adjacent works.

Candidate dates. 2015-09-23: arXiv 1509.06849v1.

Domains. General ML: graphical models (central).

Methodologies. Probabilistic inference (used); Search & optimization (used).

Concepts. Combinatorial optimization (central); Message passing (central); Graph structure & transformations (central); Constraint-aware design (central).

Keywords. blossom contraction; blossom expansion; Combinatorial optimization; Constraint-aware design; distributed algorithm; General learning & inference; Graph structure & transformations; Graphs & discrete problems; linear program; Message passing; perfect matching; Probabilistic inference; Search & optimization.

Sources. arXiv 1509.06849v1: Abstract; method formulation; experimental results and limitations.

Specific connections.

• Odd-cycle matching BP — Matching through graph transformations (documented); see r001.
• Unsupervised combinatorial optimization — Theory guides the discrete solver (interpretive); see r028.

Curated connection evidence

Matching through graph transformations

Blossom belief propagation ↔ Odd-cycle matching BP · documented.

Both adapt belief propagation to matching with graph transformations and LP structure. Blossom contraction/expansion and odd-cycle constraints yield different constructions and guarantees.

Short explanation. Both papers change a graph so belief propagation can solve a matching problem. Blossom-BP contracts and expands blossoms. Odd-cycle BP adds odd-cycle constraints. The constructions and guarantees differ.

Evidence: Blossom belief propagation: Abstract; method formulation; Odd-cycle matching BP: Abstract; method formulation.

Repairing an approximate inference engine

Odd-cycle matching BP ↔ MCMC + belief propagation · interpretive.

Odd-cycle transformations improve BP convergence for matching, while loop-based MCMC corrects BP partition estimates. Both add structure around BP, but optimization and normalization are different targets.

Short explanation. Odd-cycle BP changes a graph to improve convergence for matching. MCMC + BP samples loop corrections to estimate a partition function. The target quantities differ.

Evidence: Odd-cycle matching BP: Abstract; method formulation; MCMC + belief propagation: Abstract; method formulation.

Gauge-preserving inference

Gauged variational inference ↔ Gauged mini-bucket elimination · documented.

Both optimize gauge transformations that preserve the partition function. Gauged variational inference changes variational approximations; gauged mini-bucket combines the transformation with bounded variable elimination.

Short explanation. Both papers use gauge transformations that preserve the partition function. G-MF / G-BP changes the variational approximation. WMBE-G combines gauge transformations with bounded variable elimination.

Evidence: Gauged variational inference: Abstract; method formulation; Gauged mini-bucket elimination: Abstract; method formulation.

Bounded elimination and factor compression

Gauged mini-bucket elimination ↔ Bucket renormalization · documented.

Both control graphical-model inference complexity through bucket-based elimination. Weighted/gauged bounds and low-rank renormalization make different approximation choices.

Short explanation. Both papers use bucket elimination to control inference cost. WMBE-G uses weighted bounds and gauge transformations. MBR / GBR uses low-rank factor approximations.

Evidence: Gauged mini-bucket elimination: Abstract; method formulation; Bucket renormalization: Abstract; method formulation.

Improving partition estimates

MCMC + belief propagation ↔ Gauged variational inference · documented.

Both estimate unnormalized graphical-model partition functions relative to BP-related approximations. One samples loop corrections; the other optimizes a partition-preserving gauge.

Short explanation. Both papers estimate partition functions with belief propagation as a reference. MCMC + BP samples loop corrections. G-MF / G-BP optimizes a gauge transformation that preserves the partition function.

Evidence: MCMC + belief propagation: Abstract; method formulation; Gauged variational inference: Abstract; method formulation.

Unnormalized targets, different estimators

MCMC + belief propagation ↔ Iterated energy-based flow matching · interpretive.

Loop MCMC estimates corrections to a partition function, while iterated flow matching learns a reusable continuous sampler from energies. The common unnormalized target does not make their estimators equivalent.

Short explanation. Both papers work with unnormalized distributions. MCMC + BP estimates partition-function corrections. Iterated flow learns a reusable sampler from energies. Their estimators differ.

Evidence: MCMC + belief propagation: Abstract; method formulation; Iterated energy-based flow matching: Abstract; method formulation.

Approximation with a stochastic search partner

MCMC + belief propagation ↔ Search-guided diffusion samplers · interpretive.

BP-aware MCMC corrects a graphical-model approximation; SGDS uses MCMC exploration to train a diffusion sampler. Both pair complementary engines, but one estimates loop corrections and the other amortizes energy-space search.

Short explanation. Both papers combine an approximation with MCMC search. MCMC + BP estimates loop corrections. SGDS uses MCMC search to train a diffusion sampler.

Evidence: MCMC + belief propagation: Abstract; method formulation; Search-guided diffusion samplers: Abstract; method formulation.

Compressing useful information

Bucket renormalization ↔ Variational information distillation · interpretive.

Low-rank factor projections retain information useful for inference, while variational distillation transfers neural representations. This is a compression analogy across factors and learned features, not a shared objective.

Short explanation. Both papers reduce the size of an information representation. MBR / GBR approximates graphical-model factors. VID transfers neural features. This comparison does not imply a shared objective.

Evidence: Bucket renormalization: Abstract; method formulation; Variational information distillation: Abstract; method formulation.

Distilling an expensive teacher

Variational information distillation ↔ Planner-guided imitation · documented.

Both transfer knowledge into a reusable model. VID uses a variational information objective across networks; planner-guided imitation distills subgoal-conditioned behavior into target-goal control.

Short explanation. Both papers transfer information from a teacher into a reusable model. VID uses a variational information objective. PIG transfers behavior from a subgoal planner.

Evidence: Variational information distillation: Abstract; method formulation; Planner-guided imitation: Abstract; method formulation.

Representation transfer

Variational information distillation ↔ Co-folding representations · documented.

Both study transfer between representations with different uses or architectures. VID estimates shared information; co-folding representation analysis probes and distills ligand embeddings for standalone molecular tasks.

Short explanation. Both papers transfer information between representations. VID estimates information that two networks share. The co-folding study tests and distills ligand embeddings for molecular tasks.

Evidence: Variational information distillation: Abstract; method formulation; Co-folding representations: Abstract; method formulation.

What information should an embedding retain?

Variational information distillation ↔ scTrilemma · interpretive.

Distillation preserves useful teacher information, while scTrilemma routes biological and nuisance information across a latent embedding, prior, and decoder. Retention and deliberate invariance impose different goals.

Short explanation. Both papers control which information an embedding retains. VID transfers useful teacher information. scTrilemma separates biological and nuisance information across an embedding, a prior, and a decoder.

Evidence: Variational information distillation: Abstract; method formulation; scTrilemma: Abstract; method formulation.

Adaptive allocation of representational budget

Layer-adaptive pruning ↔ DNAChunker · interpretive.

Layer-adaptive pruning allocates weight sparsity, while DNAChunker allocates token length by genomic context. Both spend representational capacity unevenly, but parameters and sequence segmentation are different objects.

Short explanation. Both papers assign different amounts of representation capacity to different parts of an input or model. LAMP assigns weight sparsity. DNAChunker assigns token lengths from genomic context.

Evidence: Layer-adaptive pruning: Abstract; method formulation; DNAChunker: Abstract; method formulation.

Bias models expose difficult examples

Learning from Failure ↔ Learning with a biased committee · documented.

Both use deliberately bias-sensitive learners to identify bias-conflicting examples without bias labels. A single failure model and a diverse committee provide different selection and knowledge-exchange mechanisms.

Short explanation. Both papers use models that learn biases to find examples that conflict with those biases. Learning from Failure uses one model. LWBC uses several models.

Evidence: Learning from Failure: Abstract; method formulation; Learning with a biased committee: Abstract; method formulation.

Spurious-correlation mitigation

Learning with a biased committee ↔ Multi-bias robust learning · documented.

Both address learned reliance on unwanted correlations. The committee infers difficult cases without bias labels; multi-bias robust learning assumes attribute annotations and explicitly balances conflicting group objectives.

Short explanation. Both papers reduce dependence on unwanted correlations. LWBC finds difficult examples without bias labels. Multi-bias learning uses attribute labels to balance conflicting group objectives.

Evidence: Learning with a biased committee: Abstract; method formulation; Multi-bias robust learning: Abstract; method formulation.

An auxiliary model redirects training

Learning from Failure ↔ Adaptive Teachers · interpretive.

Failure-based weighting highlights bias-conflicting examples; an adaptive GFlowNet teacher seeks high-loss, poorly covered states. Both redirect training through another model, with different losses and guarantees.

Short explanation. Both papers use an auxiliary model to change which examples receive training. LfF weights examples that conflict with biases. The Adaptive Teachers method targets states with high student loss.

Evidence: Learning from Failure: Abstract; method formulation; Adaptive Teachers: Abstract; method formulation.

Invariance has competing costs

Multi-bias robust learning ↔ scTrilemma · interpretive.

Multi-bias objectives can conflict across groups, while single-cell nuisance invariance can lose biological identity or expression fidelity. Both expose tradeoffs rather than treating more invariance as universally better.

Short explanation. Both papers show that invariance can have a cost. Multi-bias learning has conflicting group objectives. scTrilemma shows conflicts between nuisance invariance, biological identity, and expression fidelity.

Evidence: Multi-bias robust learning: Abstract; method formulation; scTrilemma: Abstract; method formulation.

Learned, constrained augmentation

Difficult, not too different ↔ EPIC · documented.

DND seeks difficult text examples that preserve semantics; EPIC learns context-sensitive graph edit paths. Both learn meaningful augmentation rather than arbitrary perturbation, under different text and graph constraints.

Short explanation. Both papers learn constrained data augmentations. DND changes text while it preserves meaning. EPIC learns edit paths that respect graph context. Their constraints differ.

Evidence: Difficult, not too different: README and indexed abstract; partial review; EPIC: Abstract; method formulation.

Informative difficulty as a curriculum

Difficult, not too different ↔ Adaptive Teachers · interpretive.

DND uses low confidence with semantic preservation; Adaptive Teachers use high student loss to improve sampling coverage. Difficulty guides training in both, but the meaning of difficulty and admissible changes differs.

Short explanation. Both papers use difficult examples to guide training. DND selects text changes that reduce confidence but preserve meaning. The Adaptive Teachers method selects sampler states with high student loss.

Evidence: Difficult, not too different: README and indexed abstract; partial review; Adaptive Teachers: Abstract; method formulation.

Choose informative transformation paths

EPIC ↔ MELD · interpretive.

EPIC designs graph edit interpolation for augmentation, while MELD learns a masked-diffusion corruption schedule to avoid trajectory collisions. Both depend on intermediate graph states, not the same stochastic process.

Short explanation. Both papers choose paths through intermediate graph states. EPIC uses edit interpolation for augmentation. MELD changes diffusion corruption paths to reduce trajectory collisions. The processes differ.

Evidence: EPIC: Abstract; method formulation; MELD: Abstract; method formulation.

Change graph communication to preserve useful interactions

Non-backtracking GNNs ↔ ReBind · interpretive.

Non-backtracking updates suppress redundant message walks; ReBind adds force-motivated non-bonded edges. Their shared lesson is that connectivity and propagation rules matter, with distinct graph and molecular motivations.

Short explanation. Both papers change how graph nodes exchange information. NBA-GNN suppresses repeated message walks. ReBind adds non-bonded edges from predicted forces. Their motivations and graph rules differ.

Evidence: Non-backtracking GNNs: Abstract; method formulation; ReBind: Abstract; method formulation.

Beyond independent local node decisions

Non-backtracking GNNs ↔ Structured node diffusion · documented.

Both strengthen graph-node reasoning. Non-backtracking changes information propagation; structured node diffusion predicts labels jointly under observed-label conditions.

Short explanation. Both papers improve predictions on graph nodes. NBA-GNN changes message propagation. DPM-SNC predicts labels jointly with observed labels as conditions.

Evidence: Non-backtracking GNNs: Abstract; method formulation; Structured node diffusion: Abstract; method formulation.

Separate signal from misleading regularity

AdaPert ↔ Multi-bias robust learning · interpretive.

AdaPert focuses on perturbation-responsive genes and sparse context; multi-bias learning resists spurious attributes. This is a signal-selection parallel across biological regression and classification robustness.

Short explanation. Both papers select useful signals and reduce misleading regularities. AdaPert selects perturbation-responsive genes and sparse graph context. Multi-bias learning reduces reliance on spurious attributes.

Evidence: AdaPert: Abstract; method formulation; Multi-bias robust learning: Abstract; method formulation.

Perturbation-specific biological reasoning

AdaPert ↔ PBio-Agent / LincsQA · documented.

Both predict consequences of chemical or genetic perturbations. AdaPert predicts transcriptional responses with graph context; PBio-Agent predicts regulation directions through staged language-model reasoning.

Short explanation. Both papers predict the effects of biological perturbations. AdaPert predicts transcriptional responses with graph context. PBio-Agent predicts regulation directions through a sequence of language-model tasks.

Evidence: AdaPert: Abstract; method formulation; PBio-Agent / LincsQA: Abstract; method formulation.

Explicit reasoning context before a decision

PBio-Agent / LincsQA ↔ Causal Influence Prompting · interpretive.

PBio-Agent structures biological mechanisms and progressive evidence; Causal Influence Prompting structures actions, uncertainties, and utilities. The shared contextual reasoning pattern does not establish identified causality.

Short explanation. Both papers provide explicit context before a model makes a decision. PBio-Agent uses biological mechanisms and intermediate results. CIP uses actions, uncertainties, and utilities. This similarity does not establish causality.

Evidence: PBio-Agent / LincsQA: Abstract; method formulation; Causal Influence Prompting: Abstract; method formulation.

An interpretable edit needs a valid intervention protocol

Causal Influence Prompting ↔ Concept intervention analysis · interpretive.

Causal diagrams guide agent decisions, while concept-intervention analysis studies when editing bottlenecks helps or misleads. Both caution that an interpretable representation alone does not guarantee a valid causal intervention.

Short explanation. Both papers examine interventions in interpretable models. CIP uses causal diagrams to guide decisions. The concept-intervention study tests when changes to a bottleneck help. An interpretable representation alone does not guarantee a valid causal intervention.

Evidence: Causal Influence Prompting: Abstract; method formulation; Concept intervention analysis: Abstract; method formulation.

Learning discrete solvers and judging them fairly

Learning what to defer ↔ RL4CO · documented.

Both apply reinforcement learning to combinatorial optimization. Deferred decisions target large independent sets; RL4CO exposes modular environments and controlled routing comparisons, so solver quality depends on problem and evaluation budget.

Short explanation. Both papers apply reinforcement learning to combinatorial optimization. LwD solves independent-set problems. RL4CO provides modular routing environments and controlled comparisons. Problem choice and evaluation budget affect solver comparisons.

Evidence: Learning what to defer: Abstract; method formulation; RL4CO: Abstract; method formulation.

Constrained discrete decisions

Learning what to defer ↔ Unsupervised combinatorial optimization · documented.

Both solve structured combinatorial problems under constraints. Deferral learns the order and timing of vertex decisions; probabilistic objectives and derandomization construct feasible solutions through differentiable optimization and rounding.

Short explanation. Both papers solve discrete problems with structural constraints. LwD learns when to select vertices. UCom2 uses probabilistic objectives and derandomization to construct feasible solutions.

Evidence: Learning what to defer: Abstract; method formulation; Unsupervised combinatorial optimization: Abstract; method formulation.

Theory guides the discrete solver

Unsupervised combinatorial optimization ↔ Blossom belief propagation · interpretive.

Both connect a discrete combinatorial problem to tractable mathematical subroutines. Probabilistic rounding and matching LP/BP transformations have different scopes; no general optimality guarantee transfers between them.

Short explanation. Both papers use mathematical subroutines to solve discrete problems. UCom2 uses probabilistic rounding. Blossom-BP uses matching relaxations and graph transformations. An optimality guarantee from one method does not apply to the other.

Evidence: Unsupervised combinatorial optimization: Abstract; method formulation; Blossom belief propagation: Abstract; method formulation.

Sample budgets in learned combinatorial optimization

Symmetric replay training ↔ RL4CO · documented.

Both study reinforcement-learning solvers with explicit attention to sample efficiency or controlled budgets. Symmetric replay adds equivalent trajectories without new reward calls; RL4CO supplies a modular comparison framework.

Short explanation. Both papers control sample budgets for learned combinatorial solvers. SRT adds equivalent trajectories without new reward evaluations. RL4CO provides a framework for controlled solver comparisons.

Evidence: Symmetric replay training: Abstract; method formulation; RL4CO: Abstract; method formulation.

Reuse good candidates under expensive rewards

Symmetric replay training ↔ Genetic expert-guided learning · documented.

Both reuse successful molecular or combinatorial solutions for learning. Symmetric replay creates equivalent trajectories; genetic expert learning refines candidates by mutation/crossover and trains an apprentice from priority queues.

Short explanation. Both papers reuse successful candidates to train a model. SRT creates equivalent trajectories. GEGL uses mutation, crossover, and priority queues to improve molecular candidates.

Evidence: Symmetric replay training: Abstract; method formulation; Genetic expert-guided learning: Abstract; method formulation.

Exploration followed by local refinement

Genetic expert-guided learning ↔ Local Search GFlowNets · interpretive.

Genetic experts refine generated molecules; Local Search GFlowNets backtrack and reconstruct trajectories. Both connect exploration, refinement, and replay, but reward maximization and reward-proportional sampling are distinct objectives.

Short explanation. Both papers refine generated candidates with improvement operators and reuse the results for training. GEGL uses molecular mutation and crossover. LS-GFN uses backward-policy backtracking and forward-policy reconstruction.

Evidence: Genetic expert-guided learning: Abstract; method formulation; Local Search GFlowNets: Abstract; method formulation.

Search teaches the next model

Genetic expert-guided learning ↔ Self-improved retrosynthesis · documented.

A molecular genetic expert supplies high-reward training examples; a retrosynthesis planner supplies successful routes for reaction-model imitation. Both close a search–learning loop over different design spaces.

Short explanation. Both papers use search results to train the next model. GEGL supplies high-reward molecules. Self-improved retro supplies successful reaction routes for imitation.

Evidence: Genetic expert-guided learning: Abstract; method formulation; Self-improved retrosynthesis: Abstract; method formulation.

Amortizing planner knowledge

Self-improved retrosynthesis ↔ Planner-guided imitation · documented.

Retrosynthesis imitates successful reaction routes; goal-conditioned control imitates subgoal-planner behavior. Both train reusable policies from planning, with chemical route feasibility and control execution as different constraints.

Short explanation. Both papers train reusable policies from planner results. Self-improved retro imitates successful reaction routes. PIG transfers subgoal behavior for control. Their feasibility constraints differ.

Evidence: Self-improved retrosynthesis: Abstract; method formulation; Planner-guided imitation: Abstract; method formulation.

Subgoal structure for long-horizon control

Planner-guided imitation ↔ Adaptive-grid exploration · documented.

Both use explicit structure in goal space. Planner-guided imitation transfers subgoal behavior into a policy; breadth-first exploration records achieved and unattained grid goals to guide coverage.

Short explanation. Both papers use subgoals to structure control. PIG transfers planner behavior into a policy. BEAG records reached and unreached grid goals to guide further exploration.

Evidence: Planner-guided imitation: Abstract; method formulation; Adaptive-grid exploration: Abstract; method formulation.

Spend exploration where coverage is weak

Adaptive-grid exploration ↔ Adaptive Teachers · interpretive.

Adaptive grids prioritize poorly explored goal regions; Adaptive Teachers prioritize high-loss sampler regions. The exploration analogy spans executed control goals and amortized generative distributions.

Short explanation. Both papers direct exploration toward regions with poor coverage. BEAG targets goal regions. The Adaptive Teachers method targets sampler states with high loss. Their state spaces and objectives differ.

Evidence: Adaptive-grid exploration: Abstract; method formulation; Adaptive Teachers: Abstract; method formulation.

Multiple agents contribute different uncertainty or viewpoints

Disentangled risk-sensitive MARL ↔ INDIBATOR · interpretive.

DRIMA separates teammate and environment uncertainty in cooperative RL; INDIBATOR uses individualized debating language agents. This is a coordination analogy, not a shared multi-agent algorithm.

Short explanation. Both papers use multiple agents with distinct information. DRIMA separates uncertainty about teammates and the environment. INDIBATOR uses language agents with individual histories. The algorithms differ.

Evidence: Disentangled risk-sensitive MARL: Abstract; method formulation; INDIBATOR: Abstract; method formulation.

Retrosynthesis under available reactants

RetCL ↔ Self-improved retrosynthesis · documented.

Both model synthesis routes rather than only scoring final molecules. RetCL selects from an available reactant inventory; self-improved planning trains reaction proposals using successful multi-step routes.

Short explanation. Both papers model synthesis routes. RetCL selects reactants from an available inventory. Self-improved retro trains reaction proposals from successful multi-step routes.

Evidence: RetCL: Abstract; method formulation; Self-improved retrosynthesis: Abstract; method formulation.

Availability constrains molecular design

RetCL ↔ RxnFlow · documented.

RetCL ranks reactant combinations from an inventory; RxnFlow constructs molecules with templates and building blocks. Both encode an available synthesis space, without proving laboratory success.

Short explanation. Both papers restrict molecular design to an available synthesis space. RetCL ranks reactant combinations. RxnFlow uses reaction templates and building blocks. Neither method guarantees laboratory success.

Evidence: RetCL: Abstract; method formulation; RxnFlow: Abstract; method formulation.

Reward-proportional molecular search

RxnFlow ↔ Local Search GFlowNets · documented.

Both use GFlowNets to obtain diverse high-reward molecular objects. RxnFlow defines reaction-template actions; Local Search GFlowNets improve sampled trajectories through backtracking and reconstruction.

Short explanation. Both papers use GFlowNets to sample diverse molecules in proportion to reward. RxnFlow uses reaction-template actions. LS-GFN improves trajectories through backtracking and reconstruction.

Evidence: RxnFlow: Abstract; method formulation; Local Search GFlowNets: Abstract; method formulation.

Feasibility built into the action space

RxnFlow ↔ Spanning-tree molecular generation · documented.

RxnFlow restricts actions to reaction templates and building blocks; STGG constrains tree construction using molecular valence. Synthetic-route feasibility and graph valence are different levels of chemical admissibility.

Short explanation. Both papers restrict the action space to enforce chemical constraints. RxnFlow uses reaction templates and building blocks. STGG uses molecular valence constraints. Route feasibility and graph valence are different constraints.

Evidence: RxnFlow: Abstract; method formulation; Spanning-tree molecular generation: Conference abstract; partial review.

Compact graph construction grammars

Spanning-tree molecular generation ↔ Gap-encoded edge lists · documented.

STGG exploits molecular spanning trees and residual edges; GEEL uses gap-encoded edge lists and bandwidth. Both exploit sparsity through structured sequential encodings.

Short explanation. Both papers use sparse graph structure for sequential generation. STGG uses spanning trees and residual edges. The gap-encoding method uses gap encoding and graph bandwidth.

Evidence: Spanning-tree molecular generation: Conference abstract; partial review; Gap-encoded edge lists: Abstract; method formulation.

Hierarchical structure for graph generation

Spanning-tree molecular generation ↔ K²-tree graph generation · documented.

STGG builds a spanning tree of a molecular graph; K²-tree generation recursively partitions adjacency structure. Both use tree-based generation, but these trees represent different objects.

Short explanation. Both papers use trees for graph generation. STGG builds a molecular spanning tree. K²-tree generation divides adjacency structure recursively. The trees represent different objects.

Evidence: Spanning-tree molecular generation: Conference abstract; partial review; K²-tree graph generation: Abstract; method formulation.

Sparse graph tokenization

Gap-encoded edge lists ↔ K²-tree graph generation · documented.

Both replace dense adjacency-sequence representations with compact graph-specific tokens. Gap encoding follows an edge list; K²-trees exploit hierarchical empty submatrices.

Short explanation. Both papers replace dense adjacency sequences with compact graph tokens. The gap-encoding method encodes an edge list. HGGT encodes a hierarchy of empty and occupied submatrices.

Evidence: Gap-encoded edge lists: Abstract; method formulation; K²-tree graph generation: Abstract; method formulation.

Choose tokens that respect the object

Gap-encoded edge lists ↔ DNAChunker · interpretive.

GEEL encodes sparse graph edges; DNAChunker learns context-dependent genomic segments. Both make tokenization part of the modeling choice, without sharing a graph or genomic grammar.

Short explanation. Both papers choose tokens that match the input structure. The gap-encoding method encodes sparse graph edges. DNAChunker learns genomic segments from context. Their token grammars differ.

Evidence: Gap-encoded edge lists: Abstract; method formulation; DNAChunker: Abstract; method formulation.

Graph diffusion beyond independent variables

Wavelet graph diffusion ↔ Structured node diffusion · documented.

Wavelet diffusion couples node and edge frequency structure; structured node diffusion couples predicted labels with observed graph labels. Both exploit graph relations inside diffusion with different output types.

Short explanation. Both papers use graph relations in diffusion models. Wave-GD couples node and edge frequency structure. DPM-SNC couples predicted labels with observed labels. Their outputs differ.

Evidence: Wavelet graph diffusion: Abstract; method formulation; Structured node diffusion: Abstract; method formulation.

Separate coarse and fine spatial information

Wavelet graph diffusion ↔ Gaussian plane-wave neural operator · interpretive.

Graph wavelets preserve multiple frequency resolutions; GPWNO separates Gaussian local structure and plane-wave global structure. This is a multi-scale representation parallel across graph signals and electronic density.

Short explanation. Both papers represent spatial information at multiple scales. Wave-GD uses graph wavelets. GPWNO combines local Gaussian and global plane-wave functions. Their signals and physical meanings differ.

Evidence: Wavelet graph diffusion: Abstract; method formulation; Gaussian plane-wave neural operator: Abstract; method formulation.

Molecular representations from complementary views

HoliMol ↔ Co-folding representations · documented.

HoliMol uses fragment and geometry views of molecules; co-folding representation analysis transfers ligand information from protein–ligand models. Both examine richer molecular embeddings, with different pretraining context.

Short explanation. Both papers use complementary views for molecular representations. HoliMol uses fragment and geometry views. The co-folding study transfers ligand information from protein–ligand models. Their pretraining contexts differ.

Evidence: HoliMol: README and indexed abstract; partial review; Co-folding representations: Abstract; method formulation.

Local and global components complement one another

HoliMol ↔ Gaussian plane-wave neural operator · interpretive.

HoliMol links local fragments with whole molecules; GPWNO combines local Gaussian and global plane-wave fields. The decomposition parallel spans embeddings and a physical density operator.

Short explanation. Both papers combine local and global information. HoliMol connects molecular fragments to whole molecules. GPWNO combines local Gaussian and global plane-wave fields. The represented quantities differ.

Evidence: HoliMol: README and indexed abstract; partial review; Gaussian plane-wave neural operator: Abstract; method formulation.

Improve GFlowNet training without changing its target

LED-GFN ↔ Pessimistic backward policy · documented.

LED-GFN learns local energy decompositions; pessimistic backward policies reduce unobserved backward flow. Both change training signals or parameterization while retaining reward-proportional terminal sampling.

Short explanation. Both papers change GFlowNet training while they retain reward-proportional sampling. LED-GFN learns local energy decompositions. PBP-GFN reduces backward flow through unseen trajectories.

Evidence: LED-GFN: Abstract; method formulation; Pessimistic backward policy: Abstract; method formulation.

Backward trajectories affect discovery

Pessimistic backward policy ↔ Local Search GFlowNets · documented.

Both exploit GFlowNet backward policies. Pessimistic training changes the learned backward flow; local search explicitly backtracks sampled objects and reconstructs improved trajectories.

Short explanation. Both papers use GFlowNet backward policies. PBP-GFN changes the learned backward flow. Local Search GFN backtracks from sampled objects and reconstructs trajectories.

Evidence: Pessimistic backward policy: Abstract; method formulation; Local Search GFlowNets: Abstract; method formulation.

Coverage through additional training trajectories

Local Search GFlowNets ↔ Adaptive Teachers · documented.

Local search supplies improved nearby trajectories; Adaptive Teachers direct the student toward high-loss regions. Both augment GFlowNet training to improve coverage, with distinct exploration criteria.

Short explanation. Both papers add training trajectories to improve GFlowNet coverage. LS-GFN improves nearby trajectories. The Adaptive Teachers method selects states with high student loss. Their selection criteria differ.

Evidence: Local Search GFlowNets: Abstract; method formulation; Adaptive Teachers: Abstract; method formulation.

Dense feedback for molecular optimization

LED-GFN ↔ Co-folding representations · interpretive.

LED-GFN decomposes terminal energy into local potentials; co-folding embeddings provide representation-level signals for molecular RL. Both enrich feedback, but only the former proves the stated GFlowNet reparameterization.

Short explanation. Both papers add feedback for molecular optimization. LED-GFN splits terminal energy into local potentials. The co-folding study uses embeddings as reinforcement-learning signals. Only LED-GFN proves the stated GFlowNet reparameterization.

Evidence: LED-GFN: Abstract; method formulation; Co-folding representations: Abstract; method formulation.

Locate feedback inside a long construction

LED-GFN ↔ Latent veracity inference · interpretive.

LED-GFN supplies local signals along object construction; latent veracity inference identifies incorrect reasoning steps. Credit decomposition and latent correctness labeling are distinct mathematical tasks.

Short explanation. Both papers locate feedback within a sequence. LED-GFN supplies local signals during object construction. VS / AVI estimates which reasoning steps are incorrect. Their mathematical tasks differ.

Evidence: LED-GFN: Abstract; method formulation; Latent veracity inference: Abstract; method formulation.

Learn generators from energies

Iterated energy-based flow matching ↔ Energy-based generator matching · documented.

Both train amortized samplers without target samples using energy-derived weighted estimates. Iterated flow matching treats continuous flows; energy-based generator matching unifies flow, diffusion, and jump-process settings.

Short explanation. Both papers train reusable samplers from energies without target samples. iEFM uses continuous flows. EGM also includes diffusion and jump processes.

Evidence: Iterated energy-based flow matching: Abstract; method formulation; Energy-based generator matching: Abstract; method formulation.

Reusable sampling under expensive energy calls

Energy-based generator matching ↔ Search-guided diffusion samplers · documented.

Both learn generators for unnormalized energy distributions. Generator matching uses importance-weighted bootstrapping; SGDS separates an exploratory MCMC Searcher from an off-policy diffusion Learner.

Short explanation. Both papers learn samplers for unnormalized energy distributions. EGM uses importance-weighted bootstrap estimates. SGDS separates MCMC exploration from diffusion training with off-policy samples.

Evidence: Energy-based generator matching: Abstract; method formulation; Search-guided diffusion samplers: Abstract; method formulation.

Coverage-driven amortized sampling

Adaptive Teachers ↔ Search-guided diffusion samplers · documented.

Adaptive Teachers target poorly learned states; SGDS uses novelty-aware search and replay to expose missing modes. Both improve amortized coverage, with different training-partner objectives.

Short explanation. Both papers improve sampler coverage with a training partner. The Adaptive Teachers method targets states with high student loss. SGDS uses novelty-aware search and replay. The partner objectives differ.

Evidence: Adaptive Teachers: Abstract; method formulation; Search-guided diffusion samplers: Abstract; method formulation.

Off-policy molecular sampling

Search-guided diffusion samplers ↔ TPS-DPS · documented.

Both reuse exploratory samples to train diffusion-based samplers. SGDS targets equilibrium energy distributions; TPS-DPS targets entire transition paths, so their sampled objects and normalizations differ.

Short explanation. Both papers reuse exploratory samples to train diffusion samplers. SGDS targets equilibrium energy distributions. TPS-DPS targets transition paths. Their sampled objects and normalization terms differ.

Evidence: Search-guided diffusion samplers: Abstract; method formulation; TPS-DPS: Abstract; method formulation.

Slow dynamics and rare protein transitions

TPS-DPS ↔ BioEmu-CV · documented.

TPS-DPS learns transition-path sampling without handcrafted CVs; BioEmu-CV learns slow collective variables from a pretrained ensemble generator. Both address rare dynamical behavior through different representations.

Short explanation. Both papers study rare dynamical behavior. TPS-DPS learns transition-path sampling without predefined collective variables. BioEmu-CV learns slow collective variables from a pretrained ensemble generator.

Evidence: TPS-DPS: Abstract; method formulation; BioEmu-CV: Abstract; method formulation.

Reuse a generative model for biological computation

BioEmu-CV ↔ Riemannian MeanFlow · interpretive.

BioEmu-CV adapts an ensemble generator for slow variables; Riemannian MeanFlow learns fast manifold flow maps for design. Dynamics-aware representation and rapid generative design solve different objectives.

Short explanation. Both papers reuse generative models for biological tasks. BioEmu-CV adapts an ensemble generator to identify slow variables. RMF learns fast flow maps for design on manifolds.

Evidence: BioEmu-CV: Abstract; method formulation; Riemannian MeanFlow: Abstract; method formulation.

Generation on non-Euclidean variables

Riemannian MeanFlow ↔ MOFFlow · documented.

Both respect manifold geometry during generative modeling. Riemannian MeanFlow learns average-velocity flow maps; MOFFlow models rotations, translations, and lattices for rigid MOF blocks.

Short explanation. Both papers use manifold geometry in generation. RMF learns flow maps from average velocity. MOFFlow models rotations, translations, and lattices for rigid blocks.

Evidence: Riemannian MeanFlow: Abstract; method formulation; MOFFlow: Abstract; method formulation.

Offline design with a learned proxy

BootGen ↔ RoMA · documented.

Both optimize biological and other designs using fixed data and predictive proxies. BootGen bootstraps generated sequences with proxy labels; RoMA adapts a smooth robust proxy near optimization candidates.

Short explanation. Both papers use fixed data and learned proxies for design. BootGen trains on generated sequences with proxy labels. RoMA adapts a smooth, robust proxy near candidate designs.

Evidence: BootGen: Abstract; method formulation; RoMA: Abstract; method formulation.

A proxy improvement can miss the intended goal

RoMA ↔ Structurally diverse molecular LLMs · interpretive.

RoMA limits exploitation of model errors, while diverse molecular LLMs optimize structural diversity rather than textual diversity. Both distinguish an optimized surrogate from the desired object-level property.

Short explanation. Both papers distinguish an optimized proxy from the intended property. RoMA limits exploitation of model errors. The molecular language models target structural diversity, which differs from text diversity.

Evidence: RoMA: Abstract; method formulation; Structurally diverse molecular LLMs: Abstract; method formulation.

Pseudo-labels shape the next learner

BootGen ↔ CORE-PO · interpretive.

BootGen learns from proxy-scored generated sequences; CORE-PO learns from confident reasoning paths. Both can amplify a model-derived feedback signal, with different ranking and policy objectives.

Short explanation. Both papers train on model-derived signals. BootGen uses generated sequences with proxy scores. CORE-PO uses confident reasoning paths. Each method can reinforce errors in its own feedback.

Evidence: BootGen: Abstract; method formulation; CORE-PO: Abstract; method formulation.

Biological tokenization as a modeling decision

DNAChunker ↔ TriProRep · documented.

DNAChunker learns variable-length genomic units; TriProRep combines amino-acid, backbone, and full-atom protein tokens. Both connect representation quality to the choice of biological units.

Short explanation. Both papers choose biological tokens to improve representations. DNAChunker learns genomic units with variable lengths. TriProRep combines amino-acid, backbone, and full-atom protein tokens.

Evidence: DNAChunker: Abstract; method formulation; TriProRep: Abstract; method formulation.

Probe what structural representations retain

TriProRep ↔ Co-folding representations · documented.

TriProRep tests protein embeddings for structure prediction; co-folding analysis probes and distills ligand embeddings for standalone tasks. Both make structural representation utility an explicit empirical question.

Short explanation. Both papers test what structural embeddings can support. TriProRep tests protein structure prediction. The co-folding study tests and distills ligand embeddings for independent molecular tasks.

Evidence: TriProRep: Abstract; method formulation; Co-folding representations: Abstract; method formulation.

Sequence and structure need compatible representations

Antibody sequence–structure decoupling ↔ TriProRep · documented.

Antibody design separates sequence generation from structure prediction; TriProRep pretrains compatible protein token views for structural uses. Both study representation choices at the sequence–structure interface.

Short explanation. Both papers study the relation between sequence and structure representations. ASSD separates sequence generation from structure prediction. TriProRep learns compatible protein token views for structural tasks.

Evidence: Antibody sequence–structure decoupling: Abstract; method formulation; TriProRep: Abstract; method formulation.

Diversity does not establish biological function

Antibody sequence–structure decoupling ↔ VibeProteinBench · interpretive.

Antibody decoupling addresses token repetition and computational affinity; VibeProteinBench diagnoses language-interfaced protein workflows with computational checks. Both leave experimental function as a separate validation step.

Short explanation. Both papers leave experimental function as a separate validation step. ASSD addresses sequence repetition and computational affinity. VibeProteinBench tests protein workflows with computational checks.

Evidence: Antibody sequence–structure decoupling: Abstract; method formulation; VibeProteinBench: Abstract; method formulation.

Evaluate scientific representations across tasks

VibeProteinBench ↔ Co-folding representations · documented.

Both propose systematic evaluation beyond a single output metric. VibeProteinBench spans recognition, engineering, and generation; co-folding analysis spans probing, generation, and optimization for small molecules.

Short explanation. Both papers evaluate representations across several tasks. VibeProteinBench tests protein recognition, design, and generation. The co-folding study tests molecular embeddings through probing, generation, and optimization.

Evidence: VibeProteinBench: Abstract; method formulation; Co-folding representations: Abstract; method formulation.

Generation quality depends on the evaluation pipeline

VibeProteinBench ↔ Packora · interpretive.

Protein workflow checks separate scientific failure stages; Packora separates crystal generation from relaxation and ranking. Both show why a final success score can confound distinct components.

Short explanation. Both papers separate stages in a scientific evaluation pipeline. VibeProteinBench identifies failures in protein workflows. Packora separates crystal generation, relaxation, and ranking. A final score can hide failures in individual stages.

Evidence: VibeProteinBench: Abstract; method formulation; Packora: Abstract; method formulation.

Recover molecular structure from text

CleanMol ↔ Molecular Structural Reasoning · documented.

Both address language models’ difficulty with molecular graph structure. CleanMol teaches deterministic SMILES parsing; structural reasoning uses explicit structural sketches as intermediate supervision.

Short explanation. Both papers teach language models to recover molecular structure from text. CleanMol teaches deterministic SMILES parsing. MSR uses explicit structural sketches as intermediate supervision.

Evidence: CleanMol: Abstract; method formulation; Molecular Structural Reasoning: Abstract; method formulation.

Structural tokens expose hidden scientific constraints

CleanMol ↔ MaskGXT / HACO · interpretive.

SMILES parsing teaches graph connectivity; MaskGXT adds symmetry tokens and crystal refinement. Both strengthen structural modeling beyond generic text tokens, in different molecular and periodic geometries.

Short explanation. Both papers add structural information to scientific tokens. CleanMol teaches SMILES connectivity. MaskGXT adds symmetry tokens and crystal refinement. Molecular graphs and periodic crystals require different constraints.

Evidence: CleanMol: Abstract; method formulation; MaskGXT / HACO: Abstract; method formulation.

Intermediate molecular reasoning with checks

Molecular Structural Reasoning ↔ MT-Mol · documented.

Structural reasoning makes molecular structure explicit; MT-Mol combines stepwise proposals with RDKit-based verification and specialist agents. Both seek grounded feedback before accepting final candidates.

Short explanation. Both papers make intermediate molecular reasoning explicit. MSR uses structural sketches. MT-Mol uses specialist agents and RDKit checks to test proposals before final selection.

Evidence: Molecular Structural Reasoning: Abstract; method formulation; MT-Mol: Abstract; method formulation.

Deliberation for molecule discovery

MT-Mol ↔ INDIBATOR · documented.

Both organize proposal, critique, and selection among scientific language-model agents. MT-Mol emphasizes chemistry tools and roles; INDIBATOR adds individualized scientific and molecular histories.

Short explanation. Both papers organize proposals, critiques, and selection among molecular language agents. MT-Mol uses chemistry tools and specialist roles. INDIBATOR adds individual scientific and molecular histories.

Evidence: MT-Mol: Abstract; method formulation; INDIBATOR: Abstract; method formulation.

Ground scientific agents in domain context

INDIBATOR ↔ PBio-Agent / LincsQA · documented.

INDIBATOR uses literature and molecular histories; PBio-Agent sequences mechanistic biological tasks and propagates confident predictions. Both use structured domain context, with molecule proposals and regulation directions as different outputs.

Short explanation. Both papers give scientific agents structured domain context. INDIBATOR uses literature and molecular histories. PBio-Agent passes confident results between biological tasks. Molecule proposals and regulation directions are different outputs.

Evidence: INDIBATOR: Abstract; method formulation; PBio-Agent / LincsQA: Abstract; method formulation.

Reasoning paths matter beyond final answers

Latent veracity inference ↔ CORE-PO · documented.

Latent veracity inference estimates individual step correctness through search and amortization; CORE-PO prefers confident reasoning paths during self-training. Both inspect intermediate reasoning while relying on model-derived signals.

Short explanation. Both papers examine reasoning steps instead of only final answers. VS / AVI estimates whether individual steps are correct. CORE-PO selects confident reasoning paths for self-training. Both depend on model-derived signals.

Evidence: Latent veracity inference: Abstract; method formulation; CORE-PO: Abstract; method formulation.

Successful search supplies reusable supervision

Latent veracity inference ↔ Self-improved retrosynthesis · interpretive.

Veracity Search supplies pseudo-labels for a step-correctness model; retrosynthesis search supplies successful routes for a reaction model. Both amortize search, but model-likelihood veracity and reaction-route feasibility are different evidence.

Short explanation. Both papers use search results to train a reusable model. VS / AVI trains a model of step correctness. Self-improved retro trains a reaction model from successful routes. Their criteria for success differ.

Evidence: Latent veracity inference: Abstract; method formulation; Self-improved retrosynthesis: Abstract; method formulation.

From rigid to flexible MOF assembly

MOFFlow ↔ MOFFlow-2 · documented.

MOFFlow models rigid-block placement; MOFFlow-2 adds novel block generation and flexible torsional variables. They retain reduced structural parameterizations with different expressivity and inputs.

Short explanation. Both papers reduce the variables needed for MOF generation. MOFFlow places rigid blocks. MOFFlow-2 adds new blocks and torsion variables. Their inputs and structural flexibility differ.

Evidence: MOFFlow: Abstract; method formulation; MOFFlow-2: Abstract; method formulation.

Relax structural assumptions in MOF generation

MOFFlow-2 ↔ AtomMOF · documented.

MOFFlow-2 introduces flexible blocks and torsions; AtomMOF predicts unconstrained atomic coordinates from building-block graphs and adsorbates. Flexibility at selected variables and all-atom flexibility are different resolutions.

Short explanation. Both papers increase structural flexibility in MOF generation. MOFFlow-2 adds flexible blocks and torsions. AtomMOF predicts unconstrained atomic coordinates from block graphs and adsorbates. Their structural resolutions differ.

Evidence: MOFFlow-2: Abstract; method formulation; AtomMOF: Abstract; method formulation.

Host–guest modeling at different levels

AtomMOF ↔ MADField · documented.

AtomMOF generates MOF–adsorbate configurations; MADField predicts equilibrium adsorption density and integrated uptake. Both study adsorption, with individual structures and thermodynamic fields as different outputs.

Short explanation. Both papers model adsorption. AtomMOF generates MOF–adsorbate structures. MADField predicts equilibrium gas density and integrated uptake. Individual structures and thermodynamic fields are different outputs.

Evidence: AtomMOF: Abstract; method formulation; MADField: Abstract; method formulation.

A physical field replaces repeated microscopic computation

MADField ↔ Gaussian plane-wave neural operator · interpretive.

MADField learns gas-density fields for adsorption; GPWNO learns electronic density from local/global bases. Both amortize field prediction but have different particles, energies, and supervision.

Short explanation. Both papers predict physical density fields with reusable models. MADField predicts gas density for adsorption. GPWNO predicts electron density from local and global basis functions. Their particles, energies, and supervision differ.

Evidence: MADField: Abstract; method formulation; Gaussian plane-wave neural operator: Abstract; method formulation.

Represent spatial signals with mixed scales

Gaussian plane-wave neural operator ↔ Spherical neural fields · interpretive.

GPWNO combines local orbitals with plane waves; HNeR combines spherical grids and coordinate features. Both balance scales in a field representation, across electron density and geophysical signals.

Short explanation. Both papers combine scales in spatial representations. GPWNO combines local orbitals and plane waves. HNeR-S combines spherical grids and coordinate features. Their target signals differ.

Evidence: Gaussian plane-wave neural operator: Abstract; method formulation; Spherical neural fields: Abstract; method formulation.

A fast approximation benefits from another solver

MADField ↔ MCMC + belief propagation · interpretive.

MADField can initialize cDFT; loop-based MCMC corrects BP estimates. Both combine computational components, with initialization and stochastic correction playing different roles.

Short explanation. Both papers combine a fast approximation with another solver. MADField can initialize classical density functional theory. MCMC + BP corrects belief-propagation estimates. Initialization and stochastic correction serve different purposes.

Evidence: MADField: Abstract; method formulation; MCMC + belief propagation: Abstract; method formulation.

Joint host and adsorbate structure generation

CatFlow ↔ AtomMOF · documented.

Both use flow-based models for coupled host–guest structures. CatFlow factorizes slab primitive cells and adsorbates; AtomMOF uses all-atom MOF coordinates and interatomic-potential steering.

Short explanation. Both papers use flow models to generate a host and an adsorbate together. CatFlow uses primitive-cell slab variables. AtomMOF uses atomic MOF coordinates and interatomic-potential guidance.

Evidence: CatFlow: Abstract; method formulation; AtomMOF: Abstract; method formulation.

Factorization reduces periodic structural complexity

CatFlow ↔ MOFFlow · documented.

CatFlow factors a catalytic slab into primitive-cell variables; MOFFlow factors porous crystals into rigid building blocks. Both choose symmetry-compatible reduced variables with distinct geometry assumptions.

Short explanation. Both papers reduce periodic structures to fewer variables. CatFlow uses primitive-cell variables for catalytic slabs. MOFFlow uses rigid building blocks for porous crystals. Their geometry assumptions differ.

Evidence: CatFlow: Abstract; method formulation; MOFFlow: Abstract; method formulation.

Conditional periodic generative models

Multimodal Crystal Flow ↔ Packora · documented.

Both use flow-based generation for crystals with optional structural information. Multimodal Crystal Flow gives atom types and geometry separate flow times; Packora targets flexible molecular crystals and controlled generation/ranking comparisons.

Short explanation. Both papers use flow models to generate crystals from optional structural information. MCFlow gives atom types and geometry separate flow times. Packora targets flexible molecular crystals and controlled generation and ranking comparisons.

Evidence: Multimodal Crystal Flow: Abstract; method formulation; Packora: Abstract; method formulation.

Discrete chemistry meets continuous geometry

Multimodal Crystal Flow ↔ MOFFlow-2 · documented.

Multimodal Crystal Flow models atom types and geometry jointly; MOFFlow-2 combines discrete block descriptions with continuous flexible assembly. Both connect discrete chemical identity and periodic coordinates through different factorizations.

Short explanation. Both papers connect discrete chemical identities with continuous geometry. MCFlow models atom types and coordinates. MOFFlow-2 combines discrete block descriptions with flexible assembly. Their factorizations differ.

Evidence: Multimodal Crystal Flow: Abstract; method formulation; MOFFlow-2: Abstract; method formulation.

Composition and symmetry structure crystal generation

MaskGXT / HACO ↔ Multimodal Crystal Flow · documented.

MaskGXT uses symmetry tokens and coordinate refinement; Multimodal Crystal Flow uses composition-aware ordering and symmetry augmentation. Both respect crystal-specific structure while using different masked and flow objectives.

Short explanation. Both papers use crystal structure in generation. MaskGXT uses symmetry tokens and coordinate refinement. MCFlow uses composition-aware ordering and symmetry augmentation. Their training objectives differ.

Evidence: MaskGXT / HACO: Abstract; method formulation; Multimodal Crystal Flow: Abstract; method formulation.

A specialized search partner improves a generator

MaskGXT / HACO ↔ Genetic expert-guided learning · interpretive.

HACO explores methodological families for crystal modeling; genetic expert search improves molecular candidates and trains an apprentice. The feedback-loop parallel spans algorithm discovery and candidate discovery.

Short explanation. Both papers use specialized search to improve a generator. HACO searches for crystal-model algorithms. Genetic EL improves molecules and trains an apprentice model. Algorithm discovery and candidate discovery are different targets.

Evidence: MaskGXT / HACO: Abstract; method formulation; Genetic expert-guided learning: Abstract; method formulation.

Flexible all-atom structure generation

Packora ↔ AtomMOF · documented.

Both avoid fixing all molecular or block internal geometry during flow generation. Packora targets molecular crystals and ranking protocols; AtomMOF targets MOF–adsorbate configurations and physical steering.

Short explanation. Both papers allow internal geometry to change during flow generation. Packora generates molecular crystals and tests ranking protocols. AtomMOF generates MOF–adsorbate structures with physical guidance.

Evidence: Packora: Abstract; method formulation; AtomMOF: Abstract; method formulation.

Hamiltonian prediction versus downstream physics

QHFlow ↔ QHFlow2 · documented.

QHFlow predicts Hamiltonians through equivariant flow matching and orbital alignment; QHFlow2 evaluates SO(2)-based Hamiltonian prediction through energies and forces. Their names do not imply the same flow objective.

Short explanation. Both papers predict electronic Hamiltonians. QHFlow uses equivariant flow matching and orbital alignment. QHFlow2 tests SO(2)-based predictions through energies and forces. Their names do not imply the same flow objective.

Evidence: QHFlow: Abstract; method formulation; QHFlow2: Abstract; method formulation.

Amortize electronic-structure computation

Gaussian plane-wave neural operator ↔ QHFlow · documented.

GPWNO predicts electron density; QHFlow predicts electronic Hamiltonians and initializes SCF. Both replace expensive electronic quantities with learned predictions, but density and Hamiltonian errors have different physical consequences.

Short explanation. Both papers use learned predictions for electronic-structure computation. GPWNO predicts electron density. QHFlow predicts Hamiltonians and initializes self-consistent field calculations. Density errors and Hamiltonian errors have different physical effects.

Evidence: Gaussian plane-wave neural operator: Abstract; method formulation; QHFlow: Abstract; method formulation.

Evaluate the downstream quantity, control the pipeline

QHFlow2 ↔ Packora · interpretive.

QHFlow2 tests energy and force accuracy beyond Hamiltonian error; Packora separates generation from ranking and relaxation. Both diagnose what an intermediate metric misses, in different scientific pipelines.

Short explanation. Both papers test quantities beyond an intermediate model score. QHFlow2 tests energy and force accuracy. Packora separates generation, ranking, and relaxation. Their scientific pipelines differ.

Evidence: QHFlow2: Abstract; method formulation; Packora: Abstract; method formulation.

Taxonomy definitions

Domains

• Proteins & genomics (d_bio; major): Sequence, structure, and biological design; includes protein and nucleic-acid tasks.
• Cells & perturbations (d_cells; major): Single-cell representations and transcriptional responses to perturbations.
• Materials (d_materials; major): Crystals, porous materials, and catalysts.
• Electronic structure (d_electronic; major): Electronic Hamiltonians, electron density, and their physical observables.
• Molecules & drug discovery (d_molecules; major): Small-molecule representation, generation, optimization, and synthesis planning.
• General ML: graphical models (d_graphical; major): Belief propagation, partition-function estimation, and factor transformations for classical probabilistic graphical models.
• General ML (d_deep; major): General methods for learning, inference, sampling, graph representation, language-model reasoning, control, and robustness.
• Weather & cosmology (d_geoscience; major): Weather and cosmic-microwave-background representation on the sphere.

Methodologies

• Flow matching (m_flow; major): Continuous generative flows and learned flow maps; separate from GFlowNets.
• Diffusion models (m_diffusion; major): Denoising, score-based, and masked discrete diffusion.
• GFlowNets (m_gfn; major): Sequential reward-proportional sampling with flow-consistency training.
• Energy-based sampling (m_energy; major): Sampling from unnormalized targets or energy-defined distributions.
• Probabilistic inference (m_inference; major): Graphical-model inference, variational estimators, latent inference, and structured prediction.
• Search & optimization (m_search; major): Combinatorial, genetic, local, planning, and other explicit search mechanisms.
• Reinforcement learning (m_rl; major): Policy or value learning for control, design, or discrete optimization.
• Representation learning (m_representation; major): Learning transferable embeddings, token views, or knowledge distillation.
• Graph neural networks (m_graph; major): Learned graph message passing and graph-aware encoders or predictions.
• Language / sequence models (m_language; major): Autoregressive or masked sequence models, language-model reasoning, and scientific agents.
• Equivariant models (m_equivariant; major): Architectures or transformations respecting geometric symmetry.
• Neural fields & operators (m_fields; major): Coordinate-based fields and operators for spatial or physical signals.
• Robust learning (m_robust; major): Spurious-correlation mitigation, uncertainty-aware learning, and robust proxy design.
• Data augmentation (m_augmentation; major): Learning or constructing additional training examples through meaningful transformations.
• Compression & compact coding (m_compression; major): Sparse model parameters, graph encodings, or compact learned representations.
• Benchmarks & evaluation (m_benchmark; major): Protocols, task suites, controlled comparisons, or reusable evaluation frameworks as a contribution.

Concepts

• Adaptive resolution (c_adaptive_resolution; detail): Allocate detail at different scales or to context-dependent parts.
• Adaptive trajectories (c_adaptive_trajectories; detail): Change the length, corruption, or refinement of construction paths.
• Adsorption & host–guest modeling (c_adsorption; detail): Predict host–guest structures, gas uptake, or spatial adsorption density.
• Algorithm discovery (c_algorithm_discovery; detail): Search for and evaluate methodological ideas, including transfers across scientific domains.
• All-atom modeling (c_all_atom; detail): Model atomic details or use atom-level views; tokenization does not imply full atomic generation.
• Amortized inference / generation (c_amortization; detail): Learn a reusable predictor or sampler that replaces repeated computation.
• Auxiliary teachers (c_auxiliary_teacher; detail): Use a coupled model to select, weight, or expose useful training cases.
• Benchmarking (c_benchmarking; detail): Build or critically study evaluation protocols; routine experiments alone do not qualify.
• Causal reasoning & interventions (c_causal_reasoning; detail): Reason with causal diagrams or test concept interventions; distinguish diagrams from identified causal models.
• Combinatorial optimization (c_combinatorial; detail): Search discrete configurations under structural constraints.
• Compression (c_compression; detail): Reduce parameter, token, or factor representation size.
• Conditional generation (c_conditioning; detail): Generate or infer objects given composition, structure, labels, or other context.
• Constraint-aware design (c_constraints; detail): Enforce or model valence, inventory, feasibility, geometry, or problem constraints.
• Context selection (c_context_selection; detail): Select informative biological knowledge or reasoning context rather than consuming every possible input.
• Contrastive learning (c_contrastive; detail): Learn representation or selection by contrasting positive and negative examples.
• Credit assignment (c_credit_assignment; detail): Decompose terminal feedback into local training signals; parallel memberships denote weaker analogies.
• Cross-modal learning (c_cross_modal; detail): Connect sequence, graph, geometry, or natural-language views.
• Curriculum & training difficulty (c_curriculum; detail): Adapt example difficulty, task order, or the distribution of training cases.
• Multi-agent deliberation (c_deliberation; detail): Proposal, critique, verification, and voting among language-model agents.
• Density / field representations (c_density_fields; detail): Predict spatial fields; parallels across adsorption, electron density, and geophysical signals have different physics.
• Derandomization (c_derandomization; detail): Convert probabilistic constructions into discrete solutions through principled rounding.
• Knowledge distillation (c_distillation; detail): Transfer learned information or planner behavior into a reusable model.
• Diversity & mode coverage (c_diversity; detail): Preserve distinct modes, structural families, trajectories, or proposals.
• Energy / reward guidance (c_energy_guidance; detail): Steer a generator using a physical energy, learned potential, or terminal reward.
• Exploration–exploitation (c_exploration; detail): Balance discovering new regions with refining or exploiting promising ones.
• Factorization & decomposition (c_factorization; detail): Choose useful parts, variables, blocks, or components of an object or objective.
• Few-step generation (c_few_step; detail): Generate in one or a small number of learned flow steps.
• Genetic algorithms (c_genetic_algorithm; detail): Refine molecular candidates with mutation and crossover.
• Graph structure & transformations (c_graph_structure; detail): Exploit sparsity, edits, rewiring, graph constraints, or structural encodings.
• Learned / classical hybrids (c_hybrid_solver; detail): Couple a predictor or probabilistic approximation with a classical solver, simulation, or correction.
• Importance sampling (c_importance_sampling; detail): Estimate target expectations using weighted proposal samples.
• Information routing (c_information_routing; detail): Control what an embedding, decoder, prior, or transferred representation carries.
• Invariance & spurious factors (c_invariance; detail): Retain intended information while reducing nuisance or spurious variation.
• Local feedback (c_local_feedback; detail): Provide stepwise, structural, or intermediate signals rather than only final scores.
• Local search (c_local_search; detail): Refine candidates or trajectories in a nearby structured neighborhood.
• Long-range propagation (c_long_range; detail): Retain information across distant graph nodes.
• Manifold geometry (c_manifolds; detail): Model spherical, rotational, or other constrained geometric spaces.
• MCMC (c_mcmc; detail): Markov-chain Monte Carlo, including molecular search and loop-correction sampling.
• Message passing (c_message_passing; detail): Propagate information through graph neighborhoods or graphical-model factors.
• Discrete–continuous states (c_mixed_state; detail): Jointly model or sample discrete and continuous variables.
• Modular frameworks (c_modularity; detail): Separate reusable environments, policies, algorithms, or evaluation components.
• Multi-fidelity learning (c_multifidelity; detail): Combine computational supervision of distinct fidelities.
• Physical observables (c_physical_observables; detail): Evaluate energies, forces, or other downstream physical quantities beyond representation error.
• Physics-informed priors (c_physics_priors; detail): Use physical basis functions, potentials, or structural decompositions.
• Simulation / proxy-to-reality gap (c_proxy_gap; detail): Study or acknowledge the gap from proxy scores or simulated references to experimentally useful outcomes; not a claim of robotics sim-to-real.
• Rare events & transition paths (c_rare_events; detail): Represent infrequent transitions or slow collective motions.
• Replay & off-policy reuse (c_replay; detail): Reuse trajectories, candidates, or samples from previous policies.
• Representation alignment (c_representation_alignment; detail): Align, probe, or transfer views across molecular or protein representations.
• Sample efficiency (c_sample_efficiency; detail): Improve learning under limited data, reward calls, or executed decisions.
• Scaling studies (c_scaling; detail): Measure model or data scaling in a controlled setting.
• Search–learning feedback (c_search_learning; detail): Use search results to train a model, and learned models to guide subsequent search.
• Self-training (c_self_training; detail): Train with model-generated pseudo-labels, successful plans, or generated examples.
• Signal–noise separation (c_signal_noise; detail): Distinguish responsive or informative signals from easy, nuisance, or misleading patterns.
• Symmetry (c_symmetry; detail): Exploit geometric, trajectory, or gauge transformations with the appropriate preservation property.
• Synthesis & reactant availability (c_synthesizability; detail): Represent reaction routes or an available reactant inventory; computational feasibility is not a laboratory guarantee.
• Time scales & dynamics (c_time_scales; detail): Study slow processes, transitions, or time-lagged behavior rather than only equilibrium structures.
• Structured tokenization (c_tokenization; detail): Encode DNA, protein, graph, or crystal objects into meaningful discrete units.
• Tool-grounded reasoning (c_tool_grounding; detail): Use chemistry tools, literature, or scientific checkers to ground model outputs.
• Competing objectives (c_tradeoffs; detail): Expose or optimize competing fidelity, invariance, diversity, or robustness goals.
• Method transfer (c_transfer; detail): Reuse a methodological idea or learned information across tasks or domains.
• Uncertainty & risk (c_uncertainty; detail): Separate sources of uncertainty or optimize distinct risk sensitivities.
• Unnormalized distributions (c_unnormalized; detail): Estimate partition functions or sample targets specified without normalization constants.
• Variance reduction (c_variance_reduction; detail): Stabilize stochastic estimators using control variates, weighting, or related techniques.
• Verification & diagnostics (c_verification; detail): Check reasoning paths, structural correctness, or intervention effectiveness.