A curated collection of resources for building "AI Scientist" systems: AI that assists scientific discovery through literature intelligence, hypothesis generation, experiment planning, tool use, evaluation, and scientific communication.
- Foundational Papers & Systems
- Surveys & Overviews
- Literature Intelligence & Knowledge Extraction
- Hypothesis Generation & Ideation
- Experiment Planning & Design
- Tool-Use & Lab Automation
- Multi-Agent Architectures for Science
- Benchmarks & Evaluation
- Datasets for Scientific AI
- Scientific Communication & Reporting
- Ethics, Safety & Responsible AI for Science
- Domain: Drug Discovery & Chemistry
- Domain: Materials Science
- Domain: Genomics & Biology
- Domain: Climate & Earth Sciences
- Domain: Physics & Mathematics
- Hugging Face Models & Resources
- Agent Frameworks (General-Purpose)
- Meta-Resources & Awesome Lists
Landmark systems that define the AI Scientist paradigm.
- The AI Scientist v1 - Sakana AI, 2024. First end-to-end system for fully automated scientific discovery. Generates ideas, writes code, runs experiments, and produces LaTeX papers with automated peer review at roughly $15 per paper.
- The AI Scientist v2 - Sakana AI, 2025. Adds agentic tree search, removes human-authored templates, and produced the first fully AI-generated paper accepted at an ICLR 2025 workshop.
- Coscientist - Boiko et al., Nature 2023. GPT-4 autonomously designs, plans, and performs chemistry experiments via web search, code execution, and robotic lab control.
- ChemCrow - Bran et al., Nature Machine Intelligence 2024. An LLM augmented with 18 chemistry tools including RDKit, PubChem, retrosynthesis engines, and robotic synthesis tools.
- Google DeepMind AI Co-Scientist - Gottweis et al., 2025. A multi-agent Gemini 2.0 system using generate-debate-evolve with tournament-based Elo ranking.
- Robin - FutureHouse, 2025. An end-to-end drug discovery system that orchestrates specialized agents from literature review through experiment design.
- FunSearch - DeepMind, Nature 2024. Pairs an LLM with a systematic evaluator to search in function space and discover new mathematical constructions.
- Agent Laboratory - Schmidgall et al., 2025. A PhD, postdoc, and reviewer agent pipeline with large cost reductions versus earlier autonomous methods.
- AI-Researcher - NeurIPS 2025. A full pipeline from literature review to publication.
- Kosmos - Supports 200 parallel agent rollouts and large code-generation runs.
- Stanford Ideation Study - Si, Yang, and Hashimoto, 2024. LLM-generated research ideas were judged more novel than expert human ideas in blind evaluation.
- What's Missing in Autonomous Research? - Ren et al., 2026. A systematization of 56 autonomous-research systems that separates what a system can produce from what it can defend, and formalizes missing release-gate architecture.
- Scientific Discovery in the Age of AI - Nature, 2023. A broad review of AI breakthroughs across the research lifecycle.
- Agentic AI for Scientific Discovery - 2025. Categorizes existing systems across chemistry, biology, and materials science.
- Towards Scientific Intelligence: LLM-based Scientific Agents - 2025. A taxonomy of LLM-based scientific agent architectures and evaluation benchmarks.
Tools for searching, retrieving, synthesizing, and extracting knowledge from scientific literature.
- PaperQA2 - FutureHouse. An agentic RAG system for scientific literature search, summarization, and contradiction detection.
- OpenScholar - Allen AI and UW, Nature 2025. An open retrieval-augmented LM that searches 45M open-access papers.
- Elicit - An AI research assistant for structured literature extraction and screening.
- BGPT - REST, Python, and MCP API for scientific paper search, returning structured study evidence such as methods, sample sizes, limitations, quality scores, conflicts of interest, and data availability.
- OpenAlex - A fully open knowledge graph with hundreds of millions of works, authors, and institutions.
- Semantic Scholar API - Allen AI. A free API with 200M+ papers, citation graphs, recommendations, and AI-generated TL;DRs.
- S2ORC - Allen AI. A large open corpus of machine-readable full-text academic papers.
- ORKG - A machine-actionable platform for comparing research contributions.
- scite.ai - Smart citation analysis that classifies citations as supporting, contrasting, or mentioning.
- Connected Papers - A visual graph tool for exploring citation neighborhoods.
- CORE - An open-access full-text aggregator with hundreds of millions of metadata records.
- Galactica - Meta AI. A 120B-parameter scientific language model trained on papers, textbooks, and knowledge bases.
- SciBERT - Allen AI. BERT trained on 1.14M scientific papers with a scientific vocabulary.
- BioGPT - Microsoft. A biomedical language model with strong QA and relation-extraction performance.
- BioMedLM - Stanford CRFM. A 2.7B-parameter biomedical language model.
Systems for AI-driven hypothesis generation, novelty detection, and research ideation.
- SciMON - Wang et al., ACL 2024. Retrieves inspirations from prior work and iteratively optimizes ideas for novelty.
- ResearchAgent - Microsoft and KAIST, 2024. Defines problems, proposes methods, and designs experiments with reviewing agents.
- IDEAgent - A multi-agent research ideation system that applies quality-diversity search to evolve populations of candidate ideas, retaining diverse, high-quality directions for hypothesis generation.
- Scideator - UW and Allen AI. Decomposes papers into facets and recombines them for structured ideation.
- KG-CoI - Xiong et al., 2024. A knowledge-graph-grounded hypothesis generation system with hallucination detection.
Bayesian optimization, adaptive design, and intelligent experiment selection.
- BoTorch - Meta. A modular Bayesian optimization library on PyTorch.
- Ax Platform - Meta. A production-grade experiment management platform built on BoTorch.
- BayBE - Merck KGaA. Bayesian optimization for chemistry with built-in chemical encodings and transfer learning.
- Dragonfly - Scalable Bayesian optimization for high-dimensional and multi-fidelity problems.
- Summit - Chemical reaction optimization benchmarks and strategies.
- Olympus - Algorithm benchmarking with experimentally derived datasets.
- Phoenics - Bayesian optimization for categorical search spaces.
- GRYFFIN - Descriptor-aware optimization for categorical chemical spaces.
Frameworks connecting AI reasoning to scientific databases, code execution, and physical instruments.
- Opentrons - Open-source lab robotics with a Python API.
- ChemOS 2.0 - An open-source self-driving lab orchestration platform using SiLA2.
- Emerald Cloud Lab - Remote access to hundreds of instruments with built-in lab information systems.
- Benchling - Biotech data infrastructure with ELN, LIMS, and molecular biology tooling.
- PubChem - A massive chemistry database with programmatic access via PUG-REST.
- UniProt - Comprehensive protein sequence and functional data.
- ChEMBL - A large database of bioactive compounds with REST access.
- Materials Project - Computed materials data with API access via mp-api.
- PDB (Protein Data Bank) - Experimentally determined macromolecular structures with REST and GraphQL interfaces.
Systems where multiple specialized agents collaborate, debate, and refine scientific work.
- Agon - Sun et al., 2026. An autonomous large-scale omnidisciplinary research system built around Prompt Economy with scientist, coder, and auditor loops coordinated by prompt dispatchers.
- SciAgents - Ghafarollahi and Buehler, MIT 2024. A KG plus multi-agent pipeline for materials discovery built on AutoGen.
- AgenticSciML - 2025. Uses proposers, critics, engineers, retrievers, and evaluators for scientific ML.
- VirSci - Su et al., 2024. Uses agents with different scientific backgrounds to improve research potential.
- AgentRxiv - Schmidgall et al., 2025. A centralized preprint server where autonomous agents publish and build on each other's work.
- Multi-Agent Debate - Du et al., ICML 2024. Foundational work showing that debate between LLM instances improves reasoning.
- ChatDev - A multi-agent software development framework adaptable to scientific workflows.
- CAMEL - A role-playing agent communication framework.
Measuring scientific AI capabilities across reasoning, coding, and autonomous research.
- ScienceAgentBench - ICLR 2025. A benchmark with 102 tasks drawn from peer-reviewed scientific workflows.
- GPQA - Rein et al., 2023. Graduate-level Google-proof questions across biology, physics, and chemistry.
- SciCode - NeurIPS 2024. Research coding problems across 16 scientific subdomains.
- MLAgentBench - A benchmark suite for research-agent experimentation ability.
- SciFact - A benchmark for scientific claim verification.
- SciBench - Open-ended collegiate STEM problem solving.
- BioASQ - Biomedical information retrieval and question answering.
- IdeaBench - A benchmark for evaluating research-idea generation quality.
- IdeaGene-Bench - A benchmark for evaluating scientific lineage reasoning and lineage-grounded idea generation.
- REFUTE - Apache-2.0 benchmark for evaluating scientific critique, falsification, uncertainty calibration, overclaim detection, and evidence-grounded reasoning in LLMs.
| Dataset | Domain | Size | Path or link |
|---|---|---|---|
| GPQA | Bio/Physics/Chem | 448 questions | idavidrein/gpqa |
| SciFact | Biomedical claims | 1,400 claims | allenai/scifact |
| SciQ | Science exam QA | 13,679 questions | allenai/sciq |
| PubMedQA | Biomedical research | 1K+ labeled | qiaojin/PubMedQA |
| MMLU-Pro | STEM-heavy QA | 12K+ questions | TIGER-Lab/MMLU-Pro |
| S2ORC | Full-text papers | 81.1M papers | Allen AI S2ORC |
| The Pile | Mixed scientific text | Large | EleutherAI |
| MoleculeNet | Molecular benchmarks | Multiple datasets | DeepChem docs |
| MatBench | Materials properties | 13 tasks | GitHub |
| Open Catalyst | Catalysis/DFT | 1.3M+ relaxations | OCP project |
| QM9 | Quantum chemistry | 134K molecules | Various mirrors |
Automating the last mile from results to manuscripts, figures, and reports.
- AI Scientist Paper Module - Sakana AI. Generates LaTeX manuscripts with figures, citations, and automated review.
- Manubot - Open-source collaborative manuscripts via Markdown and GitHub-based automation.
- SciNote - An open-source ELN with manuscript-writing support.
- SciencePlots - One-line Matplotlib style sheets for publication figures.
- LLM4Papers - An AI editing agent for Overleaf documents.
- Paperpal - Academic language improvement trained on scholarly writing.
- Writefull - An academic English writing assistant with Overleaf integration.
- tikzplotlib - Converts Matplotlib figures to TikZ and PGFPlots for LaTeX.
- ScholarCopilot - TIGER-Lab. LLM trained for academic writing that drafts text with accurate, retrieval-grounded citations.
Dual-use concerns, hallucination, reproducibility, epistemic integrity, and governance.
- Dual-use capabilities of concern of biological AI models - Carter, Sandbrink et al., 2025. Proposes evaluation methods with preset risk thresholds and mitigation actions.
- AI Hallucination Taxonomy - Zhou et al., 2024. A classification of distorted information in AI-generated content.
- The ethics of using AI in scientific research - Resnik & Hosseini, 2024. Actionable recommendations for responsible AI use in science.
- AI Scientist Evaluation (ARI) - Beel, Kan & Baumgart, 2025. A critical assessment of AI Scientist v1 limitations, failure modes, and risks.
- EU Living Guidelines on Responsible Use of Generative AI in Research - Guidance on transparency, responsibility, and non-authorship for AI systems.
- UNESCO Recommendation on the Ethics of AI - A global standard with readiness and ethical impact assessment tools.
- OECD AI in Science - Policy proposals on research integrity, open science, and LLM quality standards.
- Project Rachel: Can an AI Become a Scholarly Author? - Discusses how AI scientists can publish under their own name (arXiv 2511.14819, 2025).
- SecureDNA - DNA synthesis screening to reduce biosecurity risks.
- Guardrails AI - Validation and control tooling for LLM outputs.
- FactScore - Fine-grained fact verification for model outputs.
- RDKit - Industry-standard open-source cheminformatics.
- DeepChem - Deep learning tooling for drug discovery, materials science, and quantum chemistry.
- TorchDrug - A GPU-accelerated graph-based molecular ML platform.
- Open Babel - Chemical file format conversion and molecular toolkit.
- DiffDock - MIT, 2022. Diffusion-based molecular docking.
- REINVENT 4 - AstraZeneca. De novo drug design via reinforcement learning.
- MoleculeSTM - Multi-modal molecular understanding across SMILES, graphs, and text.
- AutoDock Vina - A fast and widely used molecular docking system.
- Uni-Mol - A 3D molecular pretraining framework.
- AiZynthFinder - AstraZeneca. Template-based retrosynthetic planning via Monte Carlo tree search.
- RXN for Chemistry - IBM. Transformer-based chemical reaction prediction.
- ASKCOS - MIT's retrosynthesis and pathway planning platform.
- ChemCrow - An LLM plus chemistry-tools agent system.
- MACE - Higher-order equivariant message passing for interatomic potentials.
- CHGNet - A universal ML potential trained on Materials Project data.
- M3GNet - A graph neural network potential for the periodic table.
- Matbench Discovery - Benchmarks universal potentials on materials discovery tasks.
- Pymatgen - A comprehensive Python materials analysis library.
- JARVIS - NIST. DFT, ML, and force-field calculations for tens of thousands of materials.
- Materials Project API - Computed materials data with API access.
- CDVAE - Crystal Diffusion VAE for periodic materials generation.
- GNoME - DeepMind, Nature 2023. Predicts millions of stable crystals.
- AlphaFold - DeepMind's protein structure prediction system; AlphaFold 2 and 3 weights and code, with AlphaFold 3 extending coverage to DNA, RNA, and ligands.
- OpenFold 3 - Columbia. An open-source AlphaFold 3 reproduction.
- Protenix - ByteDance. Biomolecular structure prediction plus de novo protein-binder design.
- ESM-2 - Meta AI. Protein language modeling up to 15B parameters.
- ESM Cambrian - Next-generation ESM models with major speedups.
- ESMFold - Single-sequence protein folding in Hugging Face Transformers.
- Boltz-1 - Open-source biomolecular structure prediction.
- Chai-1 - Multi-purpose molecular structure prediction.
- RFDiffusion - Baker Lab. Protein design via diffusion models.
- Evo 2 - Arc Institute. A 40B-parameter genomic model trained on 9.3T DNA bases.
- Enformer - DeepMind. Predicts gene expression from DNA sequence across long contexts.
- Nucleotide Transformer - InstaDeep and NVIDIA. Foundation models for genomics.
- HyenaDNA - Stanford. Long-range genomic sequence modeling at single-nucleotide resolution.
- DNABERT-2 - A multi-species genome foundation model.
- scvi-tools - Deep learning tooling for single-cell omics.
- scGPT - A foundation model for single-cell biology.
- Geneformer - Christina Theodoris. A transformer pretrained on 30M single-cell transcriptomes.
- GraphCast - DeepMind, Science 2023. A fast 10-day weather model at 0.25-degree resolution.
- Aurora - Microsoft, 2025. A 1.3B-parameter atmospheric model spanning weather, chemistry, and air quality.
- Prithvi WxC - IBM and NASA. An open foundation model for weather and climate.
- FourCastNet - NVIDIA. Fourier neural operators for global weather forecasting.
- Pangu-Weather - Huawei, Nature 2023. A 3D neural weather model.
- ClimaX - Microsoft. A foundation model for weather and climate.
- WeatherBench 2 - Google Research and ECMWF. An evaluation framework for data-driven weather models.
- Climate Change AI - A community focused on the intersection of climate change and machine learning.
- PySR - Symbolic regression for interpretable scientific equations.
- LaSR - LLM-guided concept evolution for symbolic regression.
- LeanDojo - An end-to-end framework for AI theorem proving in Lean 4.
- AlphaProof - DeepMind's IMO-level mathematical reasoning system.
- AlphaGeometry 2 - DeepMind's olympiad geometry theorem prover.
- DeepXDE - Physics-informed neural networks for differential equations.
- torchdiffeq - Neural ordinary differential equations in PyTorch.
- SciML - The Julia ecosystem for scientific ML and differential equation solving.
| Model | HF Path | Domain | Notes |
|---|---|---|---|
| ESM-2 (3B) | facebook/esm2_t36_3B_UR50D |
Protein | 8M-15B parameter family |
| ESMFold | facebook/esmfold_v1 |
Protein structure | Single-sequence folding |
| ESM Cambrian | EvolutionaryScale/esmc-600m-2024-12 |
Protein | Up to 6B parameters |
| BioGPT | microsoft/biogpt |
Biomedical NLP | Strong PubMed QA signal |
| SciBERT | allenai/scibert_scivocab_uncased |
Scientific NLP | Scientific vocabulary |
| BioMedLM | stanford-crfm/BioMedLM |
Biomedical | 2.7B parameters |
| Galactica (6.7B) | facebook/galactica-6.7b |
General science | Part of a larger family |
| ChemBERTa | seyonec/ChemBERTa-zinc-base-v1 |
Molecular | SMILES-trained |
| Prithvi WxC | ibm-nasa-geospatial/Prithvi-WxC |
Weather/Climate | 2.3B parameters |
| Geneformer | ctheodoris/Geneformer |
Single-cell | 30M transcriptomes |
Not science-specific, but widely used to build scientific agent systems.
- LangChain - Foundation for many scientific LLM-agent applications.
- LangGraph - Stateful, multi-actor agent orchestration built on LangChain.
- AutoGen / AG2 - Microsoft. A multi-agent conversation framework.
- CrewAI - A role-playing multi-agent framework.
- Semantic Kernel - Microsoft. Enterprise-grade AI orchestration.
- Haystack - An end-to-end NLP and RAG framework from deepset.
- DSPy - Stanford. Programming language-model pipelines instead of hand-tuned prompting.
- Awesome-LLM-Scientific-Discovery - HKUST. A broad list of LLM resources across the scientific method.
- awesome-ai-for-science - A broad AI-for-science resource list across domains.
- LLMs-in-science - White Lab. A living bibliography of LLM scientific agents.
Thrilled to have you here.
Whether it's a quick typo fix, a fresh resource,
a doc polish, or a sweeping overhaul — every contribution helps this list grow.
Jump in and join the community — PRs of every size are welcome.