Science & maths: latest AI news

AI in science and mathematics: proofs, discoveries and papers, linked to the research itself.

2 Nov Deep-learning model reads mortality risk and lifespan trajectories from ordinary retinal photos, Nature paper says
2 Nov KAIST's CluSTER data selection cuts LLM fine-tuning time by up to 69.6% using about 31% of the data
1 Nov UK mathematicians meeting at the Royal Society warn AI labs could keep frontier maths results behind paywalls
1 Nov UNIST method trains small LLM agents to infer hidden roles, lifting Werewolf win rate from 41.6% to 65.6%
31 Oct Balestriero and LeCun prove JEPAs trained on 'good views' learn the same representation up to a rotation
31 Oct Researcher proposes 'generalized belief states' to decode and steer what sequence models represent
31 Oct World Laureates Forum opens in Shanghai; Daubechies wins maths prize, Lafforgue says LLMs solve some open problems
31 Oct Peking University team ports the ABACUS DFT code to Huawei's Kunpeng 920 and reports up to 10.7x speedups
30 Oct Epoch AI finds LLM use acknowledged in 17% of bioRxiv biology preprints, near 19% for specialised AI tools
30 Oct OpenAI will shut down Prism, its AI LaTeX workspace for scientists, on 1 January 2027
30 Oct SETI Institute and Breakthrough Listen let AI agents steer the Allen Telescope Array and process data in real time
30 Oct Preprint puts pretraining, fine-tuning and RL under one likelihood-based training algorithm
30 Oct Envision's Tianji model, entered as team Arcana, ranks first in ECMWF's AI Weather Quest for days 19 to 25
30 Oct AI analysis of 98 million bat calls makes Britain's largest woodland bat survey possible
30 Oct Collaborating AI agents choose a question, derive a result and test it experimentally on quantum spins in diamond
30 Oct Frontier GPT models offer candidate resolutions for about 60% of 200 randomly sampled open conjectures, study finds
30 Oct Ramanujan Challenge analysis finds a submission proved one of its published open number-theory conjectures
30 Oct Balestriero and LeCun prove JEPAs trained on different modalities learn the same representation up to a rotation
29 Oct YC Labs runs 50 GPT-6.1 Sol agents in Minecraft for 3 hours; they build a rationing bureaucracy
29 Oct Preprint proves the van der Waerden number w(2;3,21) equals 416 with a Lean 4 kernel-checked proof
29 Oct The Biological Computing Co. releases Mischief 1, a video model tuned with adapters derived from lab-grown neurons
28 Oct Anthropic expands its AI for Science program to 20 foundations and research institutions
27 Oct Axiom says AI agents formalized the classification of finite simple groups in Lean in 24 days
27 Oct Anthropic posts a paper claiming a subexponential-time algorithm for 3-SAT, with a Lean 4 formalization
26 Oct Researchers begin checking OpenAI and Anthropic claims that AI disproved the Exponential Time Hypothesis
23 Oct Physical Superintelligence posts AI-assisted results on five open quantum problems with Lean proofs
21 Oct About 2,000 GPT-Rosalind agents analyse a whole-brain CRISPR atlas in 32 minutes
19 Oct NVIDIA and UW's Institute for Protein Design release RFD4-Proteina as an open protein design model
19 Oct Quantinuum reports 99.9995% logical fidelity on Helios, a tenfold error-correction improvement
16 Oct Monkey recordings show the brain switching between seen and suppressed images about 15 times a second
15 Oct Google shows magic state cultivation on Willow, cutting logical error about 40-fold
11 Oct DeepMind's AlphaProof Nexus solves 9 Erdős problems, two open for over 50 years, with Lean proofs
10 Oct OpenAI publishes 722 math manuscripts produced by an unreleased internal model
6 Oct Alman and Vassilevska Williams post truly subquadratic 3SUM and subcubic APSP algorithms