Difference between revisions of "Science Agents"

From GISAXS
Jump to: navigation, search
(→‎Science Agents)
(→‎Science Agents)
 
(22 intermediate revisions by the same user not shown)
Line 79: Line 79:
 
* [https://github.com/TheBlewish/Automated-AI-Web-Researcher-Ollama Automated-AI-Web-Researcher-Ollama]
 
* [https://github.com/TheBlewish/Automated-AI-Web-Researcher-Ollama Automated-AI-Web-Researcher-Ollama]
 
* 2025-01: [https://arxiv.org/abs/2501.05366 Search-o1: Agentic Search-Enhanced Large Reasoning Models] ([https://search-o1.github.io/ project], [https://github.com/sunnynexus/Search-o1 code])
 
* 2025-01: [https://arxiv.org/abs/2501.05366 Search-o1: Agentic Search-Enhanced Large Reasoning Models] ([https://search-o1.github.io/ project], [https://github.com/sunnynexus/Search-o1 code])
 +
* 2026-02: [https://www.nature.com/articles/s41586-025-10072-4 Synthesizing scientific literature with retrieval-augmented language models] ([https://allenai.org/blog/openscholar-nature blog])
  
 
===Commercial===
 
===Commercial===
Line 88: Line 89:
 
* [https://periodic.com/ Periodic Labs]
 
* [https://periodic.com/ Periodic Labs]
 
* [https://edisonscientific.com/articles/announcing-edison-scientific Edison Scientific] (drug discovery, spinoff from [https://www.futurehouse.org/ FutureHouse])
 
* [https://edisonscientific.com/articles/announcing-edison-scientific Edison Scientific] (drug discovery, spinoff from [https://www.futurehouse.org/ FutureHouse])
 +
* 2026-03: Mirendil Inc.: advanced models to speed up R&D in scientific domains, especially biology and materials science
  
 
====Bio====
 
====Bio====
Line 130: Line 132:
 
* 2025-09: [https://www.biorxiv.org/content/10.1101/2025.09.12.675911v1 Generative design of novel bacteriophages with genome language models]
 
* 2025-09: [https://www.biorxiv.org/content/10.1101/2025.09.12.675911v1 Generative design of novel bacteriophages with genome language models]
 
* 2025-10: [https://www.science.org/doi/10.1126/science.adu8578 Strengthening nucleic acid biosecurity screening against generative protein design tools]
 
* 2025-10: [https://www.science.org/doi/10.1126/science.adu8578 Strengthening nucleic acid biosecurity screening against generative protein design tools]
 +
* 2026-01: [https://www.nature.com/articles/s41586-025-10014-0 Advancing regulatory variant effect prediction with AlphaGenome]
  
 
===Medicine===
 
===Medicine===
Line 149: Line 152:
 
* 2025-02: [https://www.biorxiv.org/content/10.1101/2025.02.06.636901v1 From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models]
 
* 2025-02: [https://www.biorxiv.org/content/10.1101/2025.02.06.636901v1 From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models]
 
* 2025-02: [https://www.goodfire.ai/blog/interpreting-evo-2 Interpreting Evo 2: Arc Institute's Next-Generation Genomic Foundation Model]
 
* 2025-02: [https://www.goodfire.ai/blog/interpreting-evo-2 Interpreting Evo 2: Arc Institute's Next-Generation Genomic Foundation Model]
 +
* 2026-01: [https://www.goodfire.ai/research/interpretability-for-alzheimers-detection# Using Interpretability to Identify a Novel Class of Alzheimer's Biomarkers]
  
 
===Uncertainty===
 
===Uncertainty===
Line 166: Line 170:
 
** 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]
 +
* 2026-02: [https://edisonscientific.com/ Edison]: [https://lab-bench.ai/ LABBench 2]
  
 
=Science Agents=
 
=Science Agents=
Line 192: Line 197:
 
* 2025-11: [https://arxiv.org/abs/2511.02824 Kosmos: An AI Scientist for Autonomous Discovery]
 
* 2025-11: [https://arxiv.org/abs/2511.02824 Kosmos: An AI Scientist for Autonomous Discovery]
 
* 2025-11: [https://arxiv.org/abs/2511.08151 SciAgent: A Unified Multi-Agent System for Generalistic Scientific Reasoning]
 
* 2025-11: [https://arxiv.org/abs/2511.08151 SciAgent: A Unified Multi-Agent System for Generalistic Scientific Reasoning]
 +
* 2026-02: [https://arxiv.org/abs/2601.23265 PaperBanana: Automating Academic Illustration for AI Scientists]
  
 
==Science Multi-Agent Setups==
 
==Science Multi-Agent Setups==
 
* 2025-01: [https://arxiv.org/abs/2501.04227 Agent Laboratory: Using LLM Agents as Research Assistants]
 
* 2025-01: [https://arxiv.org/abs/2501.04227 Agent Laboratory: Using LLM Agents as Research Assistants]
 
* 2025-04: [https://www.nature.com/articles/s41551-025-01363-2 Coordinated AI agents for advancing healthcare] ([https://www.nature.com/articles/s41551-025-01363-2.epdf?sharing_token=CIYP3J8LZE4BX31fV3WxUdRgN0jAjWel9jnR3ZoTv0O9iD-yhgqzRaz_7VASayWRePPhWDD2xFyfuOpSXbdPaOtt7oH4nfXo7telALzNwY3V1p9SxoqBEJy2OuaJ_cA35-CYQC1XgjCNTZUw46dh1KX-Dj8e7-1Vk_RlZKFLrc8%3D pdf])
 
* 2025-04: [https://www.nature.com/articles/s41551-025-01363-2 Coordinated AI agents for advancing healthcare] ([https://www.nature.com/articles/s41551-025-01363-2.epdf?sharing_token=CIYP3J8LZE4BX31fV3WxUdRgN0jAjWel9jnR3ZoTv0O9iD-yhgqzRaz_7VASayWRePPhWDD2xFyfuOpSXbdPaOtt7oH4nfXo7telALzNwY3V1p9SxoqBEJy2OuaJ_cA35-CYQC1XgjCNTZUw46dh1KX-Dj8e7-1Vk_RlZKFLrc8%3D pdf])
 +
 +
=Science Agentic Components=
 +
==Frameworks==
 +
* [https://platform.claude.com/docs/en/agent-sdk/overview Anthropic Claude Agent SKD overview]
 +
* [https://openclaw.ai/ OpenClaw]
 +
* [https://opencode.ai/ OpenCode]
 +
* [https://github.com/OpenHands/software-agent-sdk OpenHands]
 +
* [https://github.com/lamm-mit?tab=repositories LAMM: MIT Laboratory for Atomistic and Molecular Mechanics]
 +
** [https://github.com/lamm-mit/scienceclaw ScienceClaw]: Framework for autonomous scientific investigation without central coordination.
 +
** [https://infinite-lamm.vercel.app/ Infinite]: The Infinite Corridor of Scientific Discovery. Open science, powered by many — agents and humans discovering together.
 +
 +
==Skills==
 +
* 2026-03: [https://github.com/K-Dense-AI/claude-scientific-skills/tree/main?tab=readme-ov-file#use-cases Claude Scientific Skills] (list)
  
 
=AI Science Systems=
 
=AI Science Systems=
Line 206: Line 225:
 
* 2025-12: [https://arxiv.org/abs/2512.16969 Probing Scientific General Intelligence of LLMs with Scientist-Aligned Workflows]
 
* 2025-12: [https://arxiv.org/abs/2512.16969 Probing Scientific General Intelligence of LLMs with Scientist-Aligned Workflows]
 
* 2026-01: [https://www.nature.com/articles/s43588-025-00906-6 SciSciGPT: advancing human–AI collaboration in the science of science]
 
* 2026-01: [https://www.nature.com/articles/s43588-025-00906-6 SciSciGPT: advancing human–AI collaboration in the science of science]
 +
* 2026-02: [https://allenai.org/papers/autodiscovery AUTODISCOVERY: Open-ended Scientific Discovery via Bayesian Surprise] (Allen AI (Ai2) AstraLabs, [https://allenai.org/blog/autodiscovery blog], [https://autodiscovery.allen.ai/runs tools])
  
 
===Inorganic Materials Discovery===
 
===Inorganic Materials Discovery===
Line 243: Line 263:
 
** 2025-05: Retraction: [https://economics.mit.edu/news/assuring-accurate-research-record Assuring an accurate research record]
 
** 2025-05: Retraction: [https://economics.mit.edu/news/assuring-accurate-research-record Assuring an accurate research record]
 
* 2025-02: [https://arxiv.org/abs/2502.05151 Transforming Science with Large Language Models: A Survey on AI-assisted Scientific Discovery, Experimentation, Content Generation, and Evaluation]
 
* 2025-02: [https://arxiv.org/abs/2502.05151 Transforming Science with Large Language Models: A Survey on AI-assisted Scientific Discovery, Experimentation, Content Generation, and Evaluation]
 +
* 2026-02: [https://arxiv.org/abs/2602.03837 Accelerating Scientific Research with Gemini: Case Studies and Common Techniques]
  
 
=Related Tools=
 
=Related Tools=
Line 267: Line 288:
 
* 2025-11: [https://cdn.openai.com/pdf/4a25f921-e4e0-479a-9b38-5367b47e8fd0/early-science-acceleration-experiments-with-gpt-5.pdf Early science acceleration experiments with GPT-5]
 
* 2025-11: [https://cdn.openai.com/pdf/4a25f921-e4e0-479a-9b38-5367b47e8fd0/early-science-acceleration-experiments-with-gpt-5.pdf Early science acceleration experiments with GPT-5]
 
* 2025-12: [https://andymasley.substack.com/p/ai-can-obviously-create-new-knowledge AI can obviously create new knowledge - But maybe not new concepts]
 
* 2025-12: [https://andymasley.substack.com/p/ai-can-obviously-create-new-knowledge AI can obviously create new knowledge - But maybe not new concepts]
 +
==Math==
 +
* 2023-07: [https://www.nature.com/articles/s41586-023-06004-9?utm_source=chatgpt.com Faster sorting algorithms discovered using deep reinforcement learning]
 +
* 2025-06: [https://arxiv.org/abs/2506.13131 AlphaEvolve: A coding agent for scientific and algorithmic discovery]
 +
* 2025-11: [https://arxiv.org/abs/2511.02864 Mathematical exploration and discovery at scale]
 +
* 2025-11: [https://www.nature.com/articles/s41586-025-09833-y Olympiad-level formal mathematical reasoning with reinforcement learning]
 +
* 2025-12: [https://arxiv.org/abs/2512.14575 Extremal descendant integrals on moduli spaces of curves: An inequality discovered and proved in collaboration with AI]
 +
* [https://github.com/teorth/erdosproblems/wiki/AI-contributions-to-Erd%C5%91s-problems AI Solving Erdős Problems]:
 +
** 2026-01: [https://www.erdosproblems.com/728 Erdős Problem #728] and [https://www.erdosproblems.com/729 #729] solved by Aristotle using ChatGPT 5.2 Pro
 +
** 2026-01: [https://www.erdosproblems.com/forum/thread/397 Erdős Problem #397] [https://x.com/neelsomani/status/2010215162146607128?s=20 solved] by [https://neelsomani.com/ Neel Somani] using ChatGPT 5.2 Pro
 +
** 2026-01: [https://www.erdosproblems.com/205 Erdős Problem #205] solved by Aristotle using ChatGPT 5.2 Pro
 +
** 2026-01: [https://www.erdosproblems.com/forum/thread/281 Erdős Problem #281] [https://x.com/neelsomani/status/2012695714187325745?s=20 solved] by [https://neelsomani.com/ Neel Somani] using ChatGPT 5.2 Pro
 +
** 2026-01: Google DeepMind: [https://arxiv.org/abs/2601.21442 Irrationality of rapidly converging series: a problem of Erdős and Graham]
 +
*** [https://www.erdosproblems.com/1051 Erdős Problem #1051] [https://x.com/slow_developer/status/2018321002623901885?s=20 solved] by Google DeepMind Aletheia agent
 +
** 2026-01: Google DeepMind: [https://arxiv.org/abs/2601.22401 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: [https://www.erdosproblems.com/846 Erdős Problem #846]
 +
*** [https://x.com/roydanroy/status/2026804567178953048?s=20 Google DeepMind]
 +
*** [https://x.com/mehtaab_sawhney/status/2026716221933343147?s=20 Using OpenAI internal model] (paper: [https://cdn.openai.com/infinite-sets/main_single_clean3.pdf On infinite sets with no 3 on a line])
 +
* 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]
 +
* 2026-03: Donald Knuth: [https://www-cs-faculty.stanford.edu/~knuth/papers/claude-cycles.pdf A problem in Directed Hamiltonian Cycles] solved by Filip Stappers using Claude Opus 4.6
 +
* 2026-03: Google DeepMind: [https://arxiv.org/abs/2603.09172 Reinforced Generation of Combinatorial Structures: Ramsey Numbers]
 +
 +
==Physics assistance==
 +
* 2025-03: [https://arxiv.org/abs/2503.23758 Exact solution of the frustrated Potts model with next-nearest-neighbor interactions in one dimension via AI bootstrapping] ([https://www.bnl.gov/staff/wyin Weiguo Yin])
 +
* 2025-12: [https://www.sciencedirect.com/science/article/pii/S0370269325008111 Relativistic covariance and nonlinear quantum mechanics: Tomonaga-Schwinger analysis]
 +
** [https://x.com/hsu_steve/status/1996034522308026435?s=20 Steve Hsu], [https://drive.google.com/file/d/16sxJuwsHoi-fvTFbri9Bu8B9bqA6lr1H/view Theoretical Physics with Generative AI]
 +
* 2026-02: [https://arxiv.org/abs/2602.12176 Single-minus gluon tree amplitudes are nonzero] (GPT-5.2, [https://openai.com/index/new-result-theoretical-physics/ blog])
  
−
* '''Math:'''
+
==Literature exploration==
−
** 2023-07: [https://www.nature.com/articles/s41586-023-06004-9?utm_source=chatgpt.com Faster sorting algorithms discovered using deep reinforcement learning]
+
* 2025-11: [https://arxiv.org/abs/2511.02824 Kosmos: An AI Scientist for Autonomous Discovery] ([https://edisonscientific.com/ Edison])
−
** 2025-11: [https://arxiv.org/abs/2511.02864 Mathematical exploration and discovery at scale]
+
** [https://platform.edisonscientific.com/kosmos/c4bdef64-5e9b-43b9-a365-592dd1ed7587 Nucleotide metabolism in hypothermia]
−
** 2025-11: [https://www.nature.com/articles/s41586-025-09833-y Olympiad-level formal mathematical reasoning with reinforcement learning]
+
** [https://platform.edisonscientific.com/kosmos/1fdbf827-be65-4d97-9b66-bf0da600091a Determinant of perovskite solar-cell failure]
−
** 2025-12: [https://arxiv.org/abs/2512.14575 Extremal descendant integrals on moduli spaces of curves: An inequality discovered and proved in collaboration with AI]
+
** [https://platform.edisonscientific.com/kosmos/4fb3fbdb-c449-4064-9aa6-ff4ec53131d8 Log-normal connectivity in neural networks]
−
** [https://github.com/teorth/erdosproblems/wiki/AI-contributions-to-Erd%C5%91s-problems AI Solving Erdős Problems]:
+
** [https://platform.edisonscientific.com/kosmos/c6849232-5858-4634-adf5-83780afbe3db SOD2 as driver of myocardial fibrosis]
−
*** 2026-01: [https://www.erdosproblems.com/728 Erdős Problem #728] and [https://www.erdosproblems.com/729 #729] solved by Aristotle using ChatGPT 5.2 Pro
+
** [https://platform.edisonscientific.com/kosmos/abac07da-a6bb-458f-b0ba-ef08f1be617e Protective variant of SSR1 in type 2 diabetes]
−
*** 2026-01: [https://www.erdosproblems.com/forum/thread/397 Erdős Problem #397] [https://x.com/neelsomani/status/2010215162146607128?s=20 solved] by [https://neelsomani.com/ Neel Somani] using ChatGPT 5.2 Pro
+
** [https://platform.edisonscientific.com/kosmos/a770052b-2334-4bbe-b086-5149e0f03d99 Temporal ordering in Alzheimer’s disease]
−
*** 2026-01: [https://www.erdosproblems.com/205 Erdős Problem #205] solved by Aristotle using ChatGPT 5.2 Pro
+
** [https://platform.edisonscientific.com/kosmos/28c427d2-be31-48b5-b272-28d5a1e3ea5c Mechanism of neuron vulnerability in aging]
−
*** 2026-01: [https://www.erdosproblems.com/forum/thread/281 Erdős Problem #281] [https://x.com/neelsomani/status/2012695714187325745?s=20 solved] by [https://neelsomani.com/ Neel Somani] using ChatGPT 5.2 Pro
+
==Bio design==
−
** 2026-01: [https://arxiv.org/abs/2601.07222 The motivic class of the space of genus 0 maps to the flag variety]
+
* 2023-07: [https://www.nature.com/articles/s41586-023-06415-8 De novo design of protein structure and function with RFdiffusion]
−
* '''Physics assistance:'''
+
* 2025-11: [https://www.nature.com/articles/s41586-025-09721-5 Atomically accurate de novo design of antibodies with RFdiffusion]
−
** 2025-03: [https://arxiv.org/abs/2503.23758 Exact solution of the frustrated Potts model with next-nearest-neighbor interactions in one dimension via AI bootstrapping] ([https://www.bnl.gov/staff/wyin Weiguo Yin])
+
* 2025-11: [https://deepmind.google/blog/alphafold-five-years-of-impact/ AlphaFold: Five years of impact]
−
** 2025-12: [https://www.sciencedirect.com/science/article/pii/S0370269325008111 Relativistic covariance and nonlinear quantum mechanics: Tomonaga-Schwinger analysis]
+
* 2026-01: [https://www.goodfire.ai/research/interpretability-for-alzheimers-detection# Using Interpretability to Identify a Novel Class of Alzheimer's Biomarkers]
−
*** [https://x.com/hsu_steve/status/1996034522308026435?s=20 Steve Hsu], [https://drive.google.com/file/d/16sxJuwsHoi-fvTFbri9Bu8B9bqA6lr1H/view Theoretical Physics with Generative AI]
+
==Material Discovery==
−
* '''Literature exploration:'''
+
* 2023-11: [https://doi.org/10.1038/s41586-023-06735-9 Scaling deep learning for materials discovery]
−
** 2025-11: [https://arxiv.org/abs/2511.02824 Kosmos: An AI Scientist for Autonomous Discovery] ([https://edisonscientific.com/ Edison])
 
−
*** [https://platform.edisonscientific.com/kosmos/c4bdef64-5e9b-43b9-a365-592dd1ed7587 Nucleotide metabolism in hypothermia]
 
−
*** [https://platform.edisonscientific.com/kosmos/1fdbf827-be65-4d97-9b66-bf0da600091a Determinant of perovskite solar-cell failure]
 
−
*** [https://platform.edisonscientific.com/kosmos/4fb3fbdb-c449-4064-9aa6-ff4ec53131d8 Log-normal connectivity in neural networks]
 
−
*** [https://platform.edisonscientific.com/kosmos/c6849232-5858-4634-adf5-83780afbe3db SOD2 as driver of myocardial fibrosis]
 
−
*** [https://platform.edisonscientific.com/kosmos/abac07da-a6bb-458f-b0ba-ef08f1be617e Protective variant of SSR1 in type 2 diabetes]
 
−
*** [https://platform.edisonscientific.com/kosmos/a770052b-2334-4bbe-b086-5149e0f03d99 Temporal ordering in Alzheimer’s disease]
 
−
*** [https://platform.edisonscientific.com/kosmos/28c427d2-be31-48b5-b272-28d5a1e3ea5c Mechanism of neuron vulnerability in aging]
 
−
* '''Bio design:'''
 
−
** 2023-07: [https://www.nature.com/articles/s41586-023-06415-8 De novo design of protein structure and function with RFdiffusion]
 
−
** 2025-11: [https://www.nature.com/articles/s41586-025-09721-5 Atomically accurate de novo design of antibodies with RFdiffusion]
 
−
** 2025-11: [https://deepmind.google/blog/alphafold-five-years-of-impact/ AlphaFold: Five years of impact]
 
−
* '''Material Discovery:'''
 
−
** 2023-11: [https://doi.org/10.1038/s41586-023-06735-9 Scaling deep learning for materials discovery]
 
  
 
=See Also=
 
=See Also=
 
* [[AI agents]]
 
* [[AI agents]]
 
* [https://nanobot.chat/ Nanobot.chat]: Intelligent AI for the labnetwork @ mtl.mit.edu forum
 
* [https://nanobot.chat/ Nanobot.chat]: Intelligent AI for the labnetwork @ mtl.mit.edu forum

Latest revision as of 10:10, 18 March 2026

AI Use-cases for Science

Literature

LLM extract data from papers

AI finding links in literature

(Pre) Generate Articles

Explanation

Autonomous Ideation

Adapting LLMs to Science

AI/LLM Control of Scientific Instruments/Facilities

AI/ML Methods tailored to Science

Science Foundation Models

Regression (Data Fitting)

Tabular Classification/Regression

Symbolic Regression

Literature Discovery

Commercial

Bio

AI/ML Methods in Science

Imaging

Materials

Chemistry

Biology

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.

Uncertainty

Science Benchmarks

Science Agents

Reviews

Challenges

Specific

Science Multi-Agent Setups

Science Agentic Components

Frameworks

Skills

AI Science Systems

Inorganic Materials Discovery

Materials Characterization

Chemistry

Bio

Physics

LLMs Optimized for Science

Impact of AI in Science

Related Tools

Literature Search

Data Visualization

Generative

Chemistry

Science Datasets

Genuine Discoveries

Math

Physics assistance

Literature exploration

Bio design

Material Discovery

See Also

Retrieved from "http://gisaxs.com/index.php?title=Science_Agents&oldid=8737"