Research map / Large language models / Agents, code and math
LLMs for scientific discovery
Apply LLMs to drug discovery, equation discovery, and molecular reasoning with new benchmarks.
100 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Agents, code and math, keyword cluster “programs · program synthesis · programming”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
17 papers in 2023–24 and 82 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 4.7% to 6.5%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2021 · MIND dataset for diet planning and dietary healthcare with machine learning: Dataset creation using combinatorial optimization and controllable generation with domain experts (NeurIPS 2021 · 2 citations)
- 2023 · Foundation Model for Material Science (AAAI 2023 · 19 citations)
- 2024 · Domain-Agnostic Molecular Generation with Chemical Feedback (ICLR 2024 · 17 citations)
- 2025 · Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · 6 citations)
- 2026 · VCWorld: A Biological World Model for Virtual Cell Simulation (ICLR 2026 · 2 citations)
Earliest papers here
- MIND dataset for diet planning and dietary healthcare with machine learning: Dataset creation using combinatorial optimization and controllable generation with domain experts (NeurIPS 2021 · 2 citations)
- Foundation Model for Material Science (AAAI 2023 · 19 citations)
- A Dataset for Learning University STEM Courses at Scale and Generating Questions at a Human Level (AAAI 2023)
What has been tried
33 papers here report 292 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 6%, hurt 19%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Apply chemical feedback during molecule optimization | Synthetic molecules · Property scores | Helped | Domain-Agnostic Molecular Generation with Chemical Feedback (ICLR 2024 · 17 citations) | |
| add target-agnostic constraints | JAK2 · binding affinity (median) | 6.61 -> 7.57 | Helped | Generating Novel Leads for Drug Discovery Using LLMs with Logical Feedback (AAAI 2024 · 15 citations) |
| Add bilevel optimization (simulation-based continuous optimization) | Constitutive law search (a–d) and molecule design (e–h) · loss | 90.2 -> 5.2e-5 (a); 517.0 -> 2.1e-1 (b); 83.6 -> 6.0e-2 (c); 68.4 -> 1.4e-12 (d); 8.6e-1 -> 1.3e-4 (e); 9.1 -> 1.1e-1 (f | Helped | LLM and Simulation as Bilevel Optimizers: A New Paradigm to Advance Physical Scientific Discovery (ICML 2024 · 9 citations) |
| Add formulae review to structured instruction | SciBench four-dataset average, zero-shot · accuracy | 29.26 -> 30.87 | Helped | Structured Chemistry Reasoning with Large Language Models (ICML 2024 · 8 citations) |
| Replace random Graph-GA crossover with LLM-based crossover | JNK3 · Top-10 AUC | 0.553 ± 0.136 -> 0.790 ± 0.027 | Helped | Efficient Evolutionary Search Over Chemical Space with Large Language Models (ICLR 2025 · 6 citations) |
| Add the natural-language scientific problem and variable description | Oscillation 2 · NMSE | Helped | LLM-SR: Scientific Equation Discovery via Programming with Large Language Models (ICLR 2025 · 6 citations) | |
| Add GRPO fine-tuning to Evo2-1B + Qwen3-4B | KEGG-derived reasoning · F1-score | 86.14 to 93.05 | Helped | BioReason: Incentivizing Multimodal Biological Reasoning within a DNA-LLM Model (NeurIPS 2025 · 5 citations) |
| Add multi-step inspiration retrieval and composition | TOMATO-Chem · Top MS | 3.765 -> 4.020 | Helped | MOOSE-Chem: Large Language Models for Rediscovering Unseen Chemistry Scientific Hypotheses (ICLR 2025 · 5 citations) |
| Add Snorkel filtering at τ=0.5 to BARcode | Strategy @15 · Precision | 0.694 -> 0.514 | Hurt | Imitation of Life: A Search Engine for Biologically Inspired Design (AAAI 2024 · 4 citations) |
| Exclude the lowest 10% of samples using the filter step | Filter-step evaluation (Ave. Sci) · weighted accuracy | 43.57 -> 43.85 | Helped | SciInstruct: a Self-Reflective Instruction Annotated Dataset for Training Scientific Language Models (NeurIPS 2024 · 4 citations) |
| Add chain-of-thought prompting | MolPuzzle molecule understanding: FGI · F1 | 0.934±0.005 -> 0.689±0.025 | Hurt | Can LLMs Solve Molecule Puzzles? A Multimodal Benchmark for Molecular Structure Elucidation (NeurIPS 2024 · 3 citations) |
| Use sequential route format instead of tree format | USPTO Easy and USPTO-190, N=100 · solve rate (%) | 65.5 -> 91.0; 13.1 -> 64.7 | Helped | LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · 2 citations) |
What reviewers will expect
Datasets and baselines named in the 35 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- PubChem · 3 of 35 papers
- USPTO · 2 of 35 papers
- USPTO-190 · 2 of 35 papers
- Pistachio · 2 of 35 papers
- eMolecules building blocks · 2 of 35 papers
- ChEMBL · 2 of 35 papers
Baselines
- DeepSeek-R1 · 4 of 35 papers
- GPT-4o · 3 of 35 papers
- Graph-GA · 2 of 35 papers
- REINVENT · 2 of 35 papers
- LLM-SR · 2 of 35 papers
- LaSR · 2 of 35 papers
- o1 · 2 of 35 papers
- Qwen-2.5-7B · 2 of 35 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 99 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 6 papers
- Carnegie Mellon University · 5 papers
- University of Illinois, Urbana Champaign · 5 papers
- Peking University · 4 papers
- Massachusetts Institute of Technology · 3 papers
- Harvard University · 3 papers
- Zhejiang University · 3 papers
- Amazon · 3 papers
- University of California, Los Angeles · 3 papers
- Yale University · 3 papers
Use it
Download BibTeX of 22 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Mathematical and structural reasoning benchmarks (304 papers)
- Inductive and abstract reasoning evaluation (158 papers)
- LLMs for optimization and heuristics (147 papers)
- Code generation benchmarks (96 papers)
Elsewhere in the topic
- Agent instruction-following benchmarks (176 papers)
- Long-horizon LLM agents (155 papers)
- Software engineering agents (140 papers)
- LLM-guided program synthesis (93 papers)
- Execution-based code verification (90 papers)
- Retrieval-augmented formal theorem proving (62 papers)
- LLM-driven embodied and driving agents (54 papers)
- Reasoning benchmarks and platforms (47 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).