PSI says Anthropic's Fable 5.2 solved a quantum conjecture
Scott Aaronson says he does not fully follow the proof but accepts it after Lean autoformalization.
Stories per day · 30 days
3 RSS
Relatedagents4benchmark4llm-agents4benchmarks3chemistry2hypothesis-generation2metacognition2mixture-of-experts2multi-agent2openai2reinforcement-learning2robotics2ai-scientist1ai-scientists1anthropic1bandits1biomedical1biorxiv1
36 stories, newest first · one card per story
Stories tagged #scientific-discovery
Contrastive Discovery scores competing explanations for scientific agents
Contrast sets resolve 73% of large belief-shift rejections to a specific alternative and resolve 67% more questions under the same experiment budget.
Sage-Taiyi leads Taiyi-Bench scientific ideation scores
It scores 54.0 overall versus 47.9 for AI Scientist-v2, and authors report chemistry yield gains in collaborations.
BVSR limits task info during AI-scientist idea generation
Across five proposers and 12 tasks, limited information raised originality and later selection discarded half the candidates.
LLM discovery capacity linked to monofact rates
Higher monofact rates in training data increase scientific discoveries in LLMs.
Altman says OpenAI will keep releasing AI math results
He says OpenAI holds more proofs than the hundreds already released and spent on the order of hundreds of millions of dollars.
LongChang tests whether LLMs discover altered physical laws unaided
Of 12 models, only GPT-6-Astra found a law, at a 16.7% discovery rate; the other 11 scored 0%.
SciGame tests agents on real and counterfactual scientific laws
On five frontier models, recovering the exact symbolic form stays the main bottleneck.
LongChang tests whether LLMs can discover altered physical laws
Among 12 models, only GPT-6-Astra discovered laws, at a reported rate of 16.7 percent.
EmbodiedLab benchmarks robots that experiment in order to learn
Across 30 simulated science experiments, even the strongest of six frontier agents succeeds less than half the time.
Paper defines recommender regret for regularised RL estimates
Bandit and MCTS estimators are recast as recommenders; experiments cover synthetic trees, arcade games, and scientific tasks.
Test-time memory evolution beats test-time training on math discovery
TTME updates external memory instead of weights and cuts FLOPs by 82.8–84.6% and runtime by 47.7–58.4%.
Principia builds scientific principles from test-time hypotheses
The LLM framework beats knowledge, source, and experience baselines on molecular, protein, gene, and property tasks.
OpenAI agents analyze a mouse-brain CRISPR atlas in 32 minutes
The Reasoning Atlas is on GitHub and a companion paper is on bioRxiv for labs to follow up.
Sage-A1.5 is a 35B MoE agent for long-horizon scientific work
The report says the model delegates work to subagents, checks its own progress, and produces inspectable research artifacts.
Sage-A1.5 is a 35B MoE model for long-horizon agents
It adds subagent delegation, progress checks with strategy revision, and exploratory research beyond predefined questions.
CoRE learns rule routing and constrained retrieval without labeled pairs
A router and rule-constrained retriever co-adapt with reinforcement and contrastive learning on STaRK-QA.
LLM agents reproduce 77.4% of claims in SEER cancer papers
A robustness pipeline flags potentially consequential analytical choices in 9.5% of fully reproduced findings for expert review.
DiscoveryRSI recursively improves scientific hypothesis generation
Automated scores rose from 4.86 to 5.39 after two self-improvement rounds, with larger gains in falsifiability and data resolvability.
Self-building laboratories configure and run their own experimental setups
Authors report end-to-end validation on a real quantum-dot synthesis run, from one workstation to multi-robot labs.
DiscoveryRSI recursively learns to generate research hypotheses
Automated scores rose from 4.86 to 5.39 after two self-improvement rounds across four life-science domains.
SciExam for ENSO: Can AI Agents Build Climate Models?
SciExam for ENSO tests whether AI agents can build competitive climate models of El Nino from observations alone.
Can AI Agents Make Open-Ended Scientific Discovery? Evidence from Station
Station agents rediscover 62.7% of criteria from three ICLR papers, far above prior agent systems.
EvoDuet: Bilevel Co-Evolution of Web Searching and Task Solving for Scientific Discovery
EvoDuet co-evolves web queries and solutions, raising LLM discovery scores on 21 scientific tasks.