AI research with proofs and code you can inspect
Mistral’s proof model and Anthropic’s science workbench offer ways to inspect AI research. Plus table prediction and changes to European AI deadlines.
Mistral released Leanstral 1.5 on July 2, an Apache-2.0-licensed model for mathematical proofs and software verification. It works in Lean 4, where software can check proposed proofs.
Anthropic’s Claude Science workbench takes a different approach to inspection: research figures include the code, environment and interaction history that produced them. Its reviewer agents check citations, calculations and consistency between figures and code.
Mistral reports that a pipeline translating Rust into Lean tested 57 repositories and flagged 47 violated properties. Eleven represented bugs, including five previously unreported.
The computation can be substantial. Mistral's demonstration of logarithmic insertion and deletion bounds for an AVL tree, a self-balancing search structure, used 2.7 million tokens. On a set of 672 Putnam mathematics competition problems, the reported result rose from 44 solved at a 50,000-token budget to 587 at four million tokens, with up to eight attempts.
A checked proof establishes a stated property; it doesn't establish that the property captures everything the software should do.
Top Stories
Claude Science keeps the code behind a figure attached to it. Anthropic launched the scientific workbench in beta, combining research agents, scientific tools and computing resources.
Figures include their generating code, environment and interaction history. Reviewer agents check citations, calculations and whether figures agree with code.
The app works locally on macOS or Linux and with remote research infrastructure. It can submit analyses to existing computing clusters, requesting approval before accessing new resources. Local operation still sends task-relevant context to Claude.
Google introduced a model that learns from table examples at inference time. TabFM handles classification and regression without training a separate model for each dataset. Its architecture alternates attention over rows and columns, compresses rows into vectors, and processes those vectors with a transformer.
Google says pretraining used hundreds of millions of synthetic datasets. Evaluation covered 38 classification and 13 regression datasets, ranging from 700 to 150,000 samples. The article distinguishes the untuned, single-pass model from a 32-member ensemble with added features and calibration.
The EU Council approved later deadlines for high-risk AI requirements. Its AI Act revision sets December 2, 2027 for stand-alone high-risk systems and August 2, 2028 for high-risk systems embedded in products, replacing the August 2026 timetable described in the announcement.
Providers' synthetic-content transparency solutions have a separate December 2, 2026 deadline. The revision also introduces prohibitions on specified non-consensual sexual imagery. When the Council published its June 29 announcement, entry into force still required Official Journal publication and a subsequent three-day interval.
Top Links
- Sonnet 5 (anthropic.com): Anthropic's new model targets sustained coding, browser use and terminal work; the company says higher reasoning effort can match Opus 4.8 on some tasks.
- Memora (microsoft.com): Microsoft's agent-memory research preserves detailed records and retrieves them through short descriptions, contextual tags and queries that can follow related memories.
- Keeping agent-written code reviewable (blog.jonudell.net): A practitioner describes building a 42,805-line desktop application with coding agents while reading and challenging their output.
The week in AI, once a week.