AI for Science

Anthropic says Claude helped produce an end-to-end Lean formalization of Fermat’s Last Theorem in about 11 days. The achievement was not a new mathematical proof, but a large-scale translation and verification effort organized by multiple agents.

OpenJiuwen’s ScienceDiscovery treats scientific programs as evolving artifacts, repeatedly rewriting and evaluating them through tree search. Reported experiments cover oscillatory integration, special functions, code optimization, and equation discovery.

A new study turns the end-to-end Aardvark Weather system into a probabilistic forecaster by assigning separate stochastic mechanisms to observations and learned dynamics. The approach improves the mean forecast while making ensemble spread more interpretable.

BaguanHR takes a data-first approach to 0.1-degree global forecasting by using variable-wise super-resolution to expand coarse reanalysis data. Its results suggest that scaling the training set can deliver measurable forecasting gains.

Anthropic has introduced a research preview of the Model Hardware Standard, an interface designed to connect AI agents with laboratory equipment, robots, and other physical systems. The proposal focuses first on scientific workflows, where integrating hardware from different vendors can be especially time-consuming.

A new environment called Station lets agents from different model families choose research directions, run experiments, and build a shared mathematical literature without a scripted pipeline. The paper reports new results on several construction problems and releases dialogues, proofs, and verification code.

AutoResearch links research ideation with experimental execution in a two-stage autonomous research system. Its design uses cross-model review, iterative diagnosis, and independent evidence checks to reduce unreliable findings.

RIBOSPAN is a 1.61-billion-parameter bidirectional RNA foundation model pretrained with contexts of up to 10,240 nucleotides. Its evaluations suggest that native long-context training is more reliable than simply extending a short-context model at inference time.

London startup Inherent says its Faraday agent outperformed larger Anthropic and OpenAI systems on independently reproducing published research. The more important story may be its attempt to train scientific judgment, not simply generate correct answers.

A new study introduces Top-K prompting and training for single-step retrosynthesis, allowing language models to produce multiple chemically plausible proposals. Its C3LM model, trained on roughly 45.6 million verified reactions, shows strong performance on the OOD URSA-expert-2026 benchmark and complements conventional systems.

OmniScientist is an end-to-end AI scientist designed to reason directly over heterogeneous raw evidence rather than only text, labels, or precomputed features. Its pipeline links perception with ideation, experimentation, and scientific writing.

Personalized Auto-Research argues that a genuine AI co-scientist must understand a researcher’s prior work, methods, collaborators, and communities instead of producing a generic plan for a shared goal.

MathForm combines Mathlib retrieval, compiler diagnostics, and semantic-consistency feedback to refine mathematical formalizations iteratively. Its 8B model is trained on roughly 367K verified Lean 4 examples and outperforms several specialized 32B systems on the reported benchmarks.

A new FAR pipeline lets experts specify a research direction rather than a single problem, then uses literature search, model attempts, and automated triage to prioritize mathematical work. A combinatorics pilot illustrates how this design can concentrate scarce expert attention on the most promising candidates.

Large Discovery Model combines a generative model with a Bayesian non-parametric reward surrogate, allowing AI systems to propose candidates while using experimental feedback and uncertainty to guide the next search.

Researchers from DeepMind and collaborators have redesigned the optimization problem behind combination loss analysis and refined the search with machine learning and AlphaEvolve. The work improves the known upper bound on the matrix multiplication exponent from 2.371339 to 2.371177.

CADENA reconstructs a 3D mesh as an editable parametric CAD program by adding operations one at a time and checking the intermediate geometry. The work also introduces CADENA-Bench, a benchmark focused on real mechanical parts.

EMBL AI Librarian reframes Europe PMC as a natural-language evidence interface for AI agents. Instead of returning only paper lists, it plans searches, reads selected papers, and surfaces snippets that support an answer.

Anthropic reportedly used an unreleased Claude Mythos model for cryptanalysis, producing results involving HAWK and reduced-round AES. The episode is less about AI “breaking cryptography” and more about how AI may reshape expert research workflows.

DeepMind’s former AlphaFold team is no longer operating as a standalone unit, according to the report. Its members have moved into Gemini, AI coding, genomics, protein and enzyme design, fusion research, Isomorphic Labs, and in some cases Anthropic.

Anthropic says its Mythos security model helped uncover a serious weakness in HAWK, a post-quantum signature candidate under NIST review. The result does not break deployed cryptography, but it signals how AI could accelerate cryptographic research.