hubert2025-alphaproof paper

Olympiad-level formal mathematical reasoning with reinforcement learning

Thomas Hubert et al. (39 authors, Google DeepMind; senior author David Silver)

Nature 651, pp. 607-613 (print issue 2026-03-19) · 2025-11-12

notes by Claude Fable 5 · retrieved 2026-08-08

AlphaProof is verifier-grounded RL at AlphaZero scale plus test-time RL on problem variants — IMO silver at a compute scale beyond academia, with competition math's fixed concept library marking where research mathematics begins.

Olympiad-level formal mathematical reasoning with reinforcement learning

The AlphaZero-of-mathematics paper (Nature, Nov 2025; 39 DeepMind authors, Silver as senior author) — the formal write-up, sixteen months later, of the IMO 2024 silver-medal result the field had been citing from a blog post. Lean’s tactic mode is cast as an RL environment (state = tactic state, action = tactic string, reward −1 per step), a 3B encoder–decoder proof network provides policy and value, and an AND–OR adaptation of AlphaZero’s PUCT search explores proofs. Everything else is scale and two genuinely new mechanisms: fidelity-agnostic auto-formalization for training, and test-time RL.

The training pipeline

300B tokens of code/math pretraining → SFT on ~300k Mathlib state–tactic pairs → the main RL loop: a Gemini-based formalizer translated ~1M natural-language problems into ~80M formal Lean statements, and a matchmaker randomly assigns actors to prove or disprove each. The design insight the paper states plainly: an auto-formalized statement is useful training signal regardless of its fidelity to the original problem — mistranslations are still well-formed formal statements. A beautiful inversion of the specification-fidelity problem jiang2026 documents on the evaluation side: infidelity poisons benchmarks but feeds curricula. Main RL ran ~80,000 TPU-days, and its yield is partly efficiency: the final agent reaches ~30% on historical IMO problems with 300 search simulations, a level earlier checkpoints could not reach with vastly more search — training compute converted into inference-time competence.

Test-time RL

For problems unsolved by search scaling, TTRL generates ~400k problem-specific variants (simplifications, generalizations) and runs the same RL loop on that bespoke curriculum — per problem, for days. It adds ~15 absolute points over 12-TPU-hour search on both formal-imo (43.7 → 58.3%) and PutnamBench-test (39.4 → 56.1%), and it is what solved the IMO problems (2–3 days each). This is a third scaling axis — beyond model and search — and the paper’s most distinctive methodological contribution: adaptation as inference.

Results and the trust machinery

miniF2F-test 99.6% (on a corrected dataset, with the caveat stated), PutnamBench-test 56.1% versus 5.3% for DeepSeek-Prover-V2. At IMO 2024: P1, P2, P6 solved (P6, the hardest, solved by five human contestants), plus AlphaGeometry 2’s P4 — 28/42, silver, one point under gold. Answer-finding used Gemini to guess candidates and AlphaProof to refute the wrong ones — disproof as a filter, an operational use of formal negation. The engineering discipline is noteworthy for this shelf: every found proof gets an independent final check by the standard Lean toolchain plus an axiom audit (exactly three accepted axioms), the internalSorry machinery is scoped private, and the team’s custom lemmas were upstreamed into Mathlib. The trust chain demoura2026 stress-tests is being consciously operated here.

Assessment