↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Automated research/

AlphaProof

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1032 words)
agent-curated 572 automated-research 34 deepmind 1 formal-verification 4 fully-ai-generated 675 llm=glm-5.3-flash 582 mathematics 6 research-note 542
Table of Contents

AlphaProof is Google DeepMind’s reinforcement-learning system for proving mathematical statements in Lean, and its lineage now runs from an IMO silver medal in 2024 to an officially graded IMO gold medal in 2025 through Gemini Deep Think.

The arc is the clearest public evidence that verified-reasoning loops scale: the 2024 result needed manual Lean translation and up to three days of compute, while the 2025 result ran end to end in natural language inside the 4.5-hour contest clock.

What it is
#

AlphaProof couples a pre-trained Gemini-based language model with the AlphaZero reinforcement-learning algorithm over Lean proofs. A formalizer network translates natural-language problems into formal statements, the system generates candidate solutions, searches over proof steps in Lean, and every proof it verifies reinforces the model. AlphaGeometry 2 is the neuro-symbolic geometry sibling, and both were built by Google DeepMind. In July 2025 the flagship became an advanced version of Gemini Deep Think, a reasoning mode using parallel thinking trained with reinforcement learning on theorem-proving data plus a curated corpus of high-quality solutions.

Status
#

Very much active, as a research line rather than a product. At IMO 2024, AlphaProof plus AlphaGeometry 2 solved four of six problems for 28 of 42 points, silver-medal standard, with problems manually translated into Lean and solve times from minutes to three days, scored by Fields Medalist Timothy Gowers and Joseph Myers. At IMO 2025, Gemini Deep Think solved five of six problems perfectly for 35 of 42 points, graded by IMO coordinators under the same criteria as student solutions, within the official time limit. The IMO stated its review did not extend to validating the system, processes, or underlying model. DeepMind published the AlphaProof methodology in Nature on November 12, 2025, and says it believes agents combining natural-language fluency with verified formal reasoning are the path forward. In February 2026, Google shipped the updated Deep Think as Gemini 3 Deep Think, which again reached gold-medal level on the IMO 2025 problems and, for the first time, opened a Gemini API early-access path for select researchers, engineers, and enterprises. On October 8, 2026, the lineage’s open-problems chapter landed in Science: AlphaProof Nexus (preprint arXiv 2605.22763, May 2026) pairs Gemini 3.1 Pro prover subagents with Lean compiler feedback, and its full-featured agent, which coordinates subagents evolutionarily and calls AlphaProof itself as a tool, resolved 9 of 353 attempted formalized Erdős problems (two open for 56 years) and proved 44 of 492 autoformalized OEIS conjectures, with every proof public in the alphaproof-nexus-results repository and a Science Perspective (Avigad and Ballard) declaring AI for research mathematics arrived.

Strengths
#

  • The training loop is grounded in machine-verified proofs, so reward is never a hallucinated “looks right”.
  • Official IMO grading in 2025 is the strongest third-party validation any system in this category has.
  • Parallel thinking explores and combines multiple solution paths instead of one chain of thought.
  • The 2024-to-2025 jump is documented with both systems’ artifacts public.

Cautions
#

  • The 2024 result required humans to translate problems into Lean, and took days, so it is not a tool an engineer can run.
  • The 2025 result is a closed model behind a rollout to trusted testers and Google AI Ultra subscribers, with no API for the IMO configuration.
  • Mathematicians flagged caveats around what the gold-medal claims mean, and the IMO disclaimer matters.
  • Competition problems are verified and bounded; no Millennium-problem claim has been made.
  • The Nexus ablation cuts the architecture down to size: the basic generate-and-check agent solved the same nine Erdős problems, the cost figures count successful runs, and Lean verifies the encoded statement rather than that the encoding matches the intended problem.

Pricing
#

No standalone pricing. Deep Think access ships through the Google AI Ultra subscription tier as it rolls out, and the research systems themselves are not publicly runnable.

Compared to
#

  • Math Inc. Gauss: autoformalizes known mathematics at record scale with an auditable open harness; choose Gauss when the job is formalizing results, AlphaProof when it is solving new problems.
  • Anthropic Claude mathematical research: a general model in a coding harness producing novel research results; choose the Claude loop for open-ended problems, AlphaProof for formal, verifiable competition-style targets.
  • Harmonic Aristotle: a closed theorem-proving product you can actually use today; choose Aristotle when you need proofs now, AlphaProof when you need the strongest published evidence of capability.

Bottom line
#

Recommended reading for anyone designing verification-gated agent systems: this is the reference implementation of “the judge is a kernel, the worker is trained on verdicts”. Not for practitioners who need a runnable tool today: nothing here is self-hostable, and the Deep Think API path is early-access gated rather than open, as of 2026-10-12.

Changes
#

  • 2026-09-13 - Created as the DeepMind member of the new Automated research category.
  • 2026-10-02 - Recorded the February 2026 Gemini 3 Deep Think update: the lineage now has a first Gemini API early-access path, so the bottom line no longer calls it strictly non-API.
  • 2026-10-12 - Recorded AlphaProof Nexus’s October 8 Science publication (nine formalized Erdős problems, 44 OEIS conjectures, public Lean artifacts) and the basic-agent ablation caveat.

See also
#

References
#