OpenAI 用 Astra 模型攻克 10 道数学难题，数学家既兴奋又担忧

Core Highlights

OpenAI announced that its unreleased Astra model solved 10 long-standing open math problems spanning sphere packing, error-correcting codes, and the existence of non-sofic groups. The accompanying 250-plus-page paper and Lean formal verification results make this "AI doing math" not a demo but an auditable, rigorous proof. The significance is that the results can be checked by machines rather than taken on faith, which is the standard mathematicians actually require before believing a claim. That alone moves the conversation from breathless hype to substance the community can engage with on its own terms and at its own careful, skeptical pace.

It also raises the bar for what counts as a credible math result in the emerging age of formal proof assistants and machine checking.

What Happened

These problems had stalled human mathematicians for years, some for decades, because they needed insights no one had yet found through conventional techniques and brute force. Astra's answers do not merely state conclusions; they come with steps verifiable line by line by the Lean proof assistant, meaning they are machine-readable and reproducible by anyone with the tool and a little patience. By publishing the process rather than only the score, OpenAI lets the math community inspect it firsthand and challenge any weak link it can find. The paper's length and verification completeness show systematic reasoning, not a lucky guess from a language model parroting patterns it half-recognized from training data it had seen before.

OpenAI briefed selected mathematicians under embargo before the public release, hoping to pre-empt a skeptical backlash from the field.

Technical Details

Astra works close to "formal theorem proving plus LLM search": the model proposes lemmas and constructs proof sketches inside Lean, and a verifier judges correctness, rejecting anything that does not fully type-check and refusing to accept hand-waving. Sphere packing involves high-dimensional geometric optimization, error-correcting codes relate to algebraic structures, and non-sofic groups sit at the frontier of group theory. Tackling multiple subfields on one model suggests a degree of general reasoning rather than narrow pattern matching tuned to a single domain or a fixed competition format. The verification loop is precisely what gives the output its credibility and separates it from fluent but ungrounded text that looks convincing yet hides a gap a reviewer would miss.

The Lean proofs are publicly checkable, so any subtle error would be found fast, which materially raises the credibility of the claims.

Versus Competitors

Google DeepMind's AlphaProof and Meta's theorem-proving explorations also attack math but often focus on a single competition or subfield with curated setups and known answer shapes. Astra covers ten cross-subfield problems at once with full verification, standing out in volume and breadth of claimed results that reach beyond any one benchmark. Compared with models that only output natural language, binding Lean verification makes results far more trustworthy, because the proof must compile, not merely sound plausible to a reader skimming it. The bar here is mathematical certainty, not persuasive prose that could conceal a hidden gap a tired reviewer overlooks under deadline pressure.

Unlike contest-focused systems, Astra's value is in breadth across fields that rarely share techniques, notation, or even vocabulary.

Industry Impact and Use Cases

For mathematical research, AI is upgrading from "computational aid" to "collaborative prover," helping open long-stuck corners that no human team had time or intuition to chase. For education, formal proofs can become teaching and training resources that students can step through interactively instead of memorizing a dead artifact. For software and hardware verification, the same machinery can prove system correctness beyond what manual review reliably catches. What worries mathematicians is that as machines mass-produce proofs humans struggle to understand, academic evaluation and authorship will be reshaped in uncomfortable ways. Simply put, Astra is both a tireless postdoc and a shadow slowly eclipsing the human halo, and the field will have to decide how to share the credit and the scrutiny before the output outpaces our ability to read it carefully.

Journals are now debating how to credit a result whose full proof no single human can reasonably grasp from end to end. Some propose formal co-authorship for the system, while others insist human understanding remains a non-negotiable bar before any result reaches print in a respected venue.