Difference between revisions of "Science Agents"
KevinYager (talk | contribs) (→Math) |
KevinYager (talk | contribs) (→Science Benchmarks) |
||
| (3 intermediate revisions by the same user not shown) | |||
| Line 173: | Line 173: | ||
** 2024-07: [https://arxiv.org/abs/2407.09413 SPIQA: A Dataset for Multimodal Question Answering on Scientific Papers] | ** 2024-07: [https://arxiv.org/abs/2407.09413 SPIQA: A Dataset for Multimodal Question Answering on Scientific Papers] | ||
** 2024-10: [https://neurips.cc/virtual/2024/98540 FEABench: Evaluating Language Models on Real World Physics Reasoning Ability] | ** 2024-10: [https://neurips.cc/virtual/2024/98540 FEABench: Evaluating Language Models on Real World Physics Reasoning Ability] | ||
| + | * 2025-07: [https://allenai.org/blog/sciarena SciArena: A New Platform for Evaluating Foundation Models in Scientific Literature Tasks] ([https://sciarena.allen.ai/ vote], [https://huggingface.co/datasets/yale-nlp/SciArena data], [https://github.com/yale-nlp/SciArena code]) | ||
* 2026-02: [https://edisonscientific.com/ Edison]: [https://lab-bench.ai/ LABBench 2] | * 2026-02: [https://edisonscientific.com/ Edison]: [https://lab-bench.ai/ LABBench 2] | ||
| + | * 2026-04: [https://arxiv.org/abs/2604.14140 LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning] ([https://longcot.ai/ site], [https://github.com/LongHorizonReasoning/longcot code]) | ||
=Science Agents= | =Science Agents= | ||
| Line 202: | Line 204: | ||
* 2026-02: [https://arxiv.org/abs/2601.23265 PaperBanana: Automating Academic Illustration for AI Scientists] | * 2026-02: [https://arxiv.org/abs/2601.23265 PaperBanana: Automating Academic Illustration for AI Scientists] | ||
* 2026-03: [https://arxiv.org/abs/2603.20179 AI Agents Can Already Autonomously Perform Experimental High Energy Physics] | * 2026-03: [https://arxiv.org/abs/2603.20179 AI Agents Can Already Autonomously Perform Experimental High Energy Physics] | ||
| + | * 2026-04: [https://github.com/AstroPilot-AI/Denario Denario]: scientific research assistant | ||
==Science Multi-Agent Setups== | ==Science Multi-Agent Setups== | ||
| Line 315: | Line 318: | ||
** 2026-03: Three problems solved using OpenAI GPT internal model. Paper: [https://arxiv.org/pdf/2603.29961 Short Proofs in Combinatorics and Number Theory] | ** 2026-03: Three problems solved using OpenAI GPT internal model. Paper: [https://arxiv.org/pdf/2603.29961 Short Proofs in Combinatorics and Number Theory] | ||
** 2026-04: [https://www.erdosproblems.com/forum/thread/1196 Erdős Problem #1196] [https://x.com/Liam06972452/status/2044051379916882067?s=20 solved] by [https://x.com/Liam06972452 Leeham] using ChatGPT 5.4 Pro | ** 2026-04: [https://www.erdosproblems.com/forum/thread/1196 Erdős Problem #1196] [https://x.com/Liam06972452/status/2044051379916882067?s=20 solved] by [https://x.com/Liam06972452 Leeham] using ChatGPT 5.4 Pro | ||
| + | ** 2026-04: [https://www.erdosproblems.com/forum/thread/258 Erdős Problem #258] [https://x.com/prz_chojecki/status/2044129595729854493?s=20 solved] by [https://x.com/prz_chojecki Przemek Chojecki] using ChatGPT 5.4 Pro | ||
* 2026-01: [https://arxiv.org/abs/2601.07222 The motivic class of the space of genus 0 maps to the flag variety] | * 2026-01: [https://arxiv.org/abs/2601.07222 The motivic class of the space of genus 0 maps to the flag variety] | ||
* 2026-02: Google DeepMind: [https://arxiv.org/abs/2602.10177 Towards Autonomous Mathematics Research] | * 2026-02: Google DeepMind: [https://arxiv.org/abs/2602.10177 Towards Autonomous Mathematics Research] | ||
Latest revision as of 12:43, 16 April 2026
Contents
- 1 AI Use-cases for Science
- 2 Science Benchmarks
- 3 Science Agents
- 4 Science Agentic Components
- 5 AI Science Systems
- 6 Impact of AI in Science
- 7 Related Tools
- 8 Science Datasets
- 9 Genuine Discoveries
- 10 See Also
AI Use-cases for Science
Literature
- alphaXiv | Explore: Understand arXiv papers
- 2026-02: Synthesizing scientific literature with retrieval-augmented language models
LLM extract data from papers
AI finding links in literature
- 2019-07: Unsupervised word embeddings capture latent knowledge from materials science literature
- 2024-11: Large language models surpass human experts in predicting neuroscience results
(Pre) Generate Articles
- 2022-12: Re3: Generating Longer Stories With Recursive Reprompting and Revision
- 2023-03: English essays: Artificial intelligence (AI) technology in OpenAI ChatGPT application: A review of ChatGPT in writing English essay
- 2023-01: Journalism: Collaborating With ChatGPT: Considering the Implications of Generative Artificial Intelligence for Journalism and Media Education
- 2023-07: Science writing: Artificial intelligence in scientific writing: a friend or a foe?
- 2024-02: Wikipedia style: Assisting in Writing Wikipedia-like Articles From Scratch with Large Language Models
- 2024-02: LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs (code)
- 2024-08: Scientific papers: The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
- 2024-09: PaperQA2: Language Models Achieve Superhuman Synthesis of Scientific Knowledge (𝕏 post, code)
- 2025-03: Reasoning to Learn from Latent Thoughts
- 2025-03: WikiAutoGen: Towards Multi-Modal Wikipedia-Style Article Generation
- 2025-04: Sleep-time Compute: Beyond Inference Scaling at Test-time
Explanation
- 2025-02: TheoremExplainAgent: Towards Multimodal Explanations for LLM Theorem Understanding (preprint)
- 2025-04: Do Two AI Scientists Agree?
Autonomous Ideation
- 2024-04: ResearchAgent: Iterative Research Idea Generation over Scientific Literature with Large Language Models
- 2024-09: Mining Causality: AI-Assisted Search for Instrumental Variables
- 2024-12: Thinking Fast and Laterally: Multi-Agentic Approach for Reasoning about Uncertain Emerging Events
- 2024-12: LLMs can realize combinatorial creativity: generating creative ideas via LLMs for scientific research
- 2024-12: LiveIdeaBench: Evaluating LLMs' Scientific Creativity and Idea Generation with Minimal Context
- 2025-01: Hypothesis Generation for Materials Discovery and Design Using Goal-Driven and Constraint-Guided LLM Agents
- 2025-02: Agentic Deep Graph Reasoning Yields Self-Organizing Knowledge Networks
- 2025-06: Predicting Empirical AI Research Outcomes with Language Models
- 2025-06: The Ideation-Execution Gap: Execution Outcomes of LLM-Generated versus Human Research Ideas
- 2026-03: AI Can Learn Scientific Taste
Adapting LLMs to Science
- 2023-06: Domain-specific chatbots for science using embeddings
- 2024-10: Personalization of Large Language Models: A Survey
- 2024-11: Adapting While Learning: Grounding LLMs for Scientific Problems with Intelligent Tool Usage Adaptation
AI/LLM Control of Scientific Instruments/Facilities
- 2023-12: Opportunities for retrieval and tool augmented large language models in scientific facilities
- 2023-12: Virtual Scientific Companion for Synchrotron Beamlines: A Prototype
- 2023-12: Autonomous chemical research with large language models
- 2024-01: Synergizing Human Expertise and AI Efficiency with Language Model for Microscopy Operation and Automated Experiment Design
- 2024-06: From Text to Test: AI-Generated Control Software for Materials Science Instruments
- 2024-12: VISION: A Modular AI Assistant for Natural Human-Instrument Interaction at Scientific User Facilities
- 2025-01: Large language models for human-machine collaborative particle accelerator tuning through natural language
- 2025-04: Operating Robotic Laboratories with Large Language Models and Teachable Agents
AI/ML Methods tailored to Science
Science Foundation Models
- 2025-08: Intern-S1: A Scientific Multimodal Foundation Model
- 2025-11: A foundation model for atomistic materials chemistry
- 2025-11: Walrus: A Cross-Domain Foundation Model for Continuum Dynamics
- 2026-01: Deep contrastive learning enables genome-wide virtual screening
Regression (Data Fitting)
- 2024-06: Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data: training on (x,y) pairs enables inferring underlying function (define it in code, invert it, compose it)
- 2024-12: OmniPred: Language Models as Universal Regressors
Tabular Classification/Regression
Symbolic Regression
Literature Discovery
- FutureHouse
- Lumina
- Automated-AI-Web-Researcher-Ollama
- 2025-01: Search-o1: Agentic Search-Enhanced Large Reasoning Models (project, code)
- 2026-02: Synthesizing scientific literature with retrieval-augmented language models (blog)
Commercial
- Sakana AI
- Cusp AI: Materials/AI
- Lila AI: Life sciences
- Radical AI: Material simulation/design
- Autoscience (Carl)
- Periodic Labs
- Edison Scientific (drug discovery, spinoff from FutureHouse)
- 2026-03: Mirendil Inc.: advanced models to speed up R&D in scientific domains, especially biology and materials science
Bio
AI/ML Methods in Science
Imaging
- 2025-05: Behind the Noise: Conformal Quantile Regression Reveals Emergent Representations (blog: Behind the Noise)
Materials
- 2024-12: Crystal structure generation with autoregressive large language modeling
- 2025-03: All-atom Diffusion Transformers: Unified generative modelling of molecules and materials
- 2022-11: Training-Free Active Learning Framework in Materials Science with Large Language Models
Chemistry
- 2025-01: Large language models for reticular chemistry
- 2025-02: Image-based generation for molecule design with SketchMol
- 2025-02: Large language models for scientific discovery in molecular property prediction
- 2025-03: Vant AI Neo-1: atomistic foundation model (small molecules, proteins, etc.)
- 2025-04: Compositional Flows for 3D Molecule and Synthesis Pathway Co-design
- 2025-07: General purpose models for the chemical sciences
- 2025-11: ChemTorch: A Deep Learning Framework for Benchmarking and Developing Chemical Reaction Property Prediction Models
Biology
- 2018: AlphaFold
- 2021-07: AlphaFold 2
- 2024-05: AlphaFold 3
- 2023-03: Evolutionary-scale prediction of atomic-level protein structure with a language model (ESMFold)
- 2023-11: Illuminating protein space with a programmable generative model
- 2024-11: Sequence modeling and design from molecular to genome scale with Evo (Evo)
- 2025-01: Targeting protein–ligand neosurfaces with a generalizable deep learning tool (Chroma)
- 2025-01: Simulating 500 million years of evolution with a language model (ESM 3 model)
- 2025-02: Genome modeling and design across all domains of life with Evo 2
- 2025-02: Exploring the structural changes driving protein function with BioEmu-1
- 2025-02: Protein Large Language Models: A Comprehensive Survey
- 2025-03: Vant AI Neo-1: atomistic foundation model (small molecules, proteins, etc.)
- 2025-03: Lyra: An Efficient and Expressive Subquadratic Architecture for Modeling Biological Sequences
- 2025-08: RosettaFold 3: Accelerating Biomolecular Modeling with AtomWorks and RF3
- 2025-09: Generative design of novel bacteriophages with genome language models
- 2025-10: Strengthening nucleic acid biosecurity screening against generative protein design tools
- 2026-01: Advancing regulatory variant effect prediction with AlphaGenome
Medicine
See: AI_Agents#Medicine
Successes
AI/ML Methods co-opted for Science
Mechanistic Interpretability
Train large model on science data. Then apply mechanistic interpretability (e.g. sparse autoencoders, SAE) to the feature/activation space.
- Mechanistic interpretability for protein language models (visualizer, code, SAE)
- Markov Bio: Through a Glass Darkly: Mechanistic Interpretability as the Bridge to End-to-End Biology (quick description, background info on recent bio progress)
- 2023-01: Tracr: Compiled Transformers as a Laboratory for Interpretability (code)
- 2024-10: An X-Ray Is Worth 15 Features: Sparse Autoencoders for Interpretable Radiology Report Generation
- 2024-12: Towards scientific discovery with dictionary learning: Extracting biological concepts from microscopy foundation models
- 2024-12: InterPLM: Discovering Interpretable Features in Protein Language Models via Sparse Autoencoders
- 2025-01: Insights on Galaxy Evolution from Interpretable Sparse Feature Networks
- 2025-02: From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models
- 2025-02: Interpreting Evo 2: Arc Institute's Next-Generation Genomic Foundation Model
- 2026-01: Using Interpretability to Identify a Novel Class of Alzheimer's Biomarkers
Uncertainty
- 2024-10: entropix: Entropy Based Sampling and Parallel CoT Decoding
- 2024-10: Taming Overconfidence in LLMs: Reward Calibration in RLHF
Science Benchmarks
- 2024-07: SciCode: A Research Coding Benchmark Curated by Scientists (project)
- 2024-11: AidanBench: Evaluating Novel Idea Generation on Open-Ended Questions (code)
- 2024-12: LiveIdeaBench: Evaluating LLMs' Scientific Creativity and Idea Generation with Minimal Context
- 2025-01: Humanity's Last Exam
- ScienceAgentBench
- 2025-02: EAIRA: Establishing a Methodology for Evaluating AI Models as Scientific Research Assistants
- 2025-03: BixBench: Novel hypotheses (accept/reject)
- 2025-04: Google: Evaluating progress of LLMs on scientific problem-solving
- 2025-07: SciArena: A New Platform for Evaluating Foundation Models in Scientific Literature Tasks (vote, data, code)
- 2026-02: Edison: LABBench 2
- 2026-04: LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning (site, code)
Science Agents
Reviews
- 2024-10: Empowering biomedical discovery with AI agents
- 2025-01: A review of large language models and autonomous agents in chemistry (github)
- 2025-07: AI4Research: A Survey of Artificial Intelligence for Scientific Research
- 2025-08: From AI for Science to Agentic Science: A Survey on Autonomous Scientific Discovery
Challenges
Specific
- 2024-01-13: ORGANA: A Robotic Assistant for Automated Chemistry Experimentation and Characterization (video)
- 2024-06-19: LLMatDesign: Autonomous Materials Discovery with Large Language Models
- 2024-08-12: Sakana AI: AI Scientist; The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery (code)
- 2024-09-09: SciAgents: Automating scientific discovery through multi-agent intelligent graph reasoning (code)
- 2024-09-11: PaperQA2: Language Models Achieve Superhuman Synthesis of Scientific Knowledge (𝕏 post, code)
- 2024-10-17: Rapid and Automated Alloy Design with Graph Neural Network-Powered LLM-Driven Multi-Agent Systems
- 2024-10-28: Large Language Model-Guided Prediction Toward Quantum Materials Synthesis
- 2024-12-06: The Virtual Lab: AI Agents Design New SARS-CoV-2 Nanobodies with Experimental Validation (writeup: Virtual lab powered by ‘AI scientists’ super-charges biomedical research: Could human–AI collaborations be the future of interdisciplinary studies?)
- 2024-12-30: Aviary: training language agents on challenging scientific tasks
- See also: AI Agents > Deep Research
- 2025-04-08: Sakana: The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search (code)
- 2025-07: DREAMS: Density Functional Theory Based Research Engine for Agentic Materials Simulation
- 2025-11: Kosmos: An AI Scientist for Autonomous Discovery
- 2025-11: SciAgent: A Unified Multi-Agent System for Generalistic Scientific Reasoning
- 2026-02: PaperBanana: Automating Academic Illustration for AI Scientists
- 2026-03: AI Agents Can Already Autonomously Perform Experimental High Energy Physics
- 2026-04: Denario: scientific research assistant
Science Multi-Agent Setups
- 2025-01: Agent Laboratory: Using LLM Agents as Research Assistants
- 2025-04: Coordinated AI agents for advancing healthcare (pdf)
Science Agentic Components
Frameworks
- Anthropic Claude Agent SKD overview
- OpenClaw
- OpenCode
- OpenHands
- LAMM: MIT Laboratory for Atomistic and Molecular Mechanics
- ScienceClaw: Framework for autonomous scientific investigation without central coordination.
- Infinite: The Infinite Corridor of Scientific Discovery. Open science, powered by many — agents and humans discovering together.
Personalities
Skills
- 2026-03: Claude Scientific Skills (list)
AI Science Systems
- 2025-01: Dolphin: Closed-loop Open-ended Auto-research through Thinking, Practice, and Feedback
- 2025-01: Hypothesis Generation for Materials Discovery and Design Using Goal-Driven and Constraint-Guided LLM Agents
- 2025-02: Towards an AI co-scientist (Google blog post: Accelerating scientific breakthroughs with an AI co-scientist)
- 2025-06: The Discovery Engine
- 2025-07: Benchmarking the Discovery Engine (blog)
- 2025-07: Autonomous Scientific Discovery Through Hierarchical AI Scientist Systems
- 2025-12: Probing Scientific General Intelligence of LLMs with Scientist-Aligned Workflows
- 2026-01: SciSciGPT: advancing human–AI collaboration in the science of science
- 2026-02: AUTODISCOVERY: Open-ended Scientific Discovery via Bayesian Surprise (Allen AI (Ai2) AstraLabs, blog, tools)
Inorganic Materials Discovery
- 2023-11: Scaling deep learning for materials discovery
- 2023-11: An autonomous laboratory for the accelerated synthesis of novel materials
- 2024-09: HoneyComb: A Flexible LLM-Based Agent System for Materials Science
- 2024-10: Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models (code, datasets, checkpoints, blogpost)
- 2025-01: A generative model for inorganic materials design
- 2025-04: System of Agentic AI for the Discovery of Metal-Organic Frameworks
- 2025-05: The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models
Materials Characterization
Chemistry
- 2023-12: Autonomous chemical research with large language models (Coscientist)
- 2024-09: PNNL ChemAIst V0.2
- 2024-11: An automatic end-to-end chemical synthesis development platform powered by large language models
- 2025-06: Training a Scientific Reasoning Model for Chemistry
- 2025-06: ChemGraph: An Agentic Framework for Computational Chemistry Workflows (code)
Bio
Physics
- 2025-12: PhysMaster: Building an Autonomous AI Physicist for Theoretical and Computational Physics Research
LLMs Optimized for Science
- 2022-11: Galactica: A Large Language Model for Science
- 2024-12: Crystal structure generation with autoregressive large language modeling
- 2025-02: MatterChat: A Multi-Modal LLM for Material Science
- 2025-03: OmniScience: A Domain-Specialized LLM for Scientific Reasoning and Discovery
- 2025-03: Google TxGemma (2B, 9B, 27B): drug development
Impact of AI in Science
- 2024-11:
Artificial Intelligence, Scientific Discovery, and Product Innovation- 2025-05: Retraction: Assuring an accurate research record
- 2025-02: Transforming Science with Large Language Models: A Survey on AI-assisted Scientific Discovery, Experimentation, Content Generation, and Evaluation
- 2026-02: Accelerating Scientific Research with Gemini: Case Studies and Common Techniques
Related Tools
Literature Search
Data Visualization
- 2024-10: Microsoft Data Formulator: Create Rich Visualization with AI iteratively (video, code)
- Julius AI: Analyze your data with computational AI
Generative
- 2025-03: StarVector 1B, 8B: text or image to SVG
Chemistry
Science Datasets
- Google Dataset Search
- Awesome Materials & Chemistry Datasets
- NIST Jarvis (simulations)
Genuine Discoveries
- 2025-11: Early science acceleration experiments with GPT-5
- 2025-12: AI can obviously create new knowledge - But maybe not new concepts
Math
- 2023-07: Faster sorting algorithms discovered using deep reinforcement learning
- 2025-06: AlphaEvolve: A coding agent for scientific and algorithmic discovery
- 2025-11: Mathematical exploration and discovery at scale
- 2025-11: Olympiad-level formal mathematical reasoning with reinforcement learning
- 2025-12: Extremal descendant integrals on moduli spaces of curves: An inequality discovered and proved in collaboration with AI
- AI Solving Erdős Problems:
- 2026-01: Erdős Problem #728 and #729 solved by Aristotle using ChatGPT 5.2 Pro
- 2026-01: Erdős Problem #397 solved by Neel Somani using ChatGPT 5.2 Pro
- 2026-01: Erdős Problem #205 solved by Aristotle using ChatGPT 5.2 Pro
- 2026-01: Erdős Problem #281 solved by Neel Somani using ChatGPT 5.2 Pro
- 2026-01: Google DeepMind: Irrationality of rapidly converging series: a problem of Erdős and Graham
- Erdős Problem #1051 solved by Google DeepMind Aletheia agent
- 2026-01: Google DeepMind: Semi-Autonomous Mathematics Discovery with Gemini: A Case Study on the Erdős Problems
- Attempted 700 problems, solved 13 open Erdős problems: 5 novel autonomous solutions, 8 through existing literature.
- 2026-02: Erdős Problem #846
- 2026-03: Three problems solved using OpenAI GPT internal model. Paper: Short Proofs in Combinatorics and Number Theory
- 2026-04: Erdős Problem #1196 solved by Leeham using ChatGPT 5.4 Pro
- 2026-04: Erdős Problem #258 solved by Przemek Chojecki using ChatGPT 5.4 Pro
- 2026-01: The motivic class of the space of genus 0 maps to the flag variety
- 2026-02: Google DeepMind: Towards Autonomous Mathematics Research
- 2026-03: Donald Knuth: A problem in Directed Hamiltonian Cycles solved by Filip Stappers using Claude Opus 4.6
- 2026-03: Google DeepMind: Reinforced Generation of Combinatorial Structures: Ramsey Numbers
- 2026-03: FrontierMath problem: "A Ramsey-style Problem on Hypergraphs" solved by Kevin Barreto and Liam Price using GPT-5.4 Pro
Physics assistance
- 2025-03: Exact solution of the frustrated Potts model with next-nearest-neighbor interactions in one dimension via AI bootstrapping (Weiguo Yin)
- 2025-12: Relativistic covariance and nonlinear quantum mechanics: Tomonaga-Schwinger analysis
- 2026-02: Single-minus gluon tree amplitudes are nonzero (GPT-5.2, blog)
Literature exploration
Bio design
- 2023-07: De novo design of protein structure and function with RFdiffusion
- 2025-11: Atomically accurate de novo design of antibodies with RFdiffusion
- 2025-11: AlphaFold: Five years of impact
- 2026-01: Using Interpretability to Identify a Novel Class of Alzheimer's Biomarkers
Material Discovery
See Also
- AI agents
- Nanobot.chat: Intelligent AI for the labnetwork @ mtl.mit.edu forum
