Classy aggregates publicly available research; we did not run Bolzano ourselves.
What changed
Researchers posted Bolzano, an open source multi agent proof search system, on arXiv on 13 October 2026 (identifier 2610.09769). The stack pairs parallel prover agents with a verifier and keeps a human readable research state. After manual runs on expert picked problems produced eight proofs checked by domain experts, the team ran Bolzano without problem specific human guidance on about 3,800 open problems extracted from four paper sets. The run reportedly solved about 200 open problems. One batch used papers accepted to STOC 2026; the authors say Bolzano answered four questions raised in those papers, confirmed by the original authors.
Why it matters
Theory and systems teams already use language models for lemma sketches and counterexample search. Bolzano is a step toward batch resolution of stated open problems with a verifier in the loop, not just autocomplete for known proof templates. If the author confirmations hold, procurement and safety reviewers will ask whether internal math and protocol teams need comparable harnesses before trusting model output in design reviews.
Who is affected
Applied researchers in algorithms and cryptography, formal methods leads, and AI platform owners who sell coding or reasoning tools to science customers. Competitive intelligence teams tracking Google DeepMind, OpenAI, and open source math agents should treat Bolzano as a reference architecture for parallel search plus verification.
What to do next
Assign one engineer to reproduce a single STOC 2026 confirmation cited in the paper on your own hardware, logging verifier failures. If you cannot reproduce, keep Bolzano off production critical paths but watch the repo for tooling you can borrow.
What to watch
Independent replication of the four STOC 2026 author confirmations and any public release of the Bolzano codebase named in the arXiv entry. The next signal is a dated replication note from a lab outside the author team.
Photo: Pexels contributor / Pexels
Photo: Pexels contributor / Pexels
Sources
- Primary — arXiv, Bolzano: From Expert Guided Proof Search to Automated Open Problem Solving (13 October 2026). Defines the multi agent architecture, dataset scale, and STOC 2026 author confirmed results.
- Secondary — AI News Brief, Bolzano announcement index (13 October 2026). Same arXiv metadata and abstract for cross check.
Related in AI
Continue reading