DeepSeek AI Unleashes Open-Source Theorem Prover, Shattering Accuracy Records with 88.9% Score

Breakthrough in Automated Mathematical Reasoning

DeepSeek AI today unveiled DeepSeek-Prover-V2, an open-source large language model that achieves a record 88.9% pass rate on the MiniF2F benchmark for formal theorem proving. The 671-billion-parameter model also cracked 49 out of 658 problems from the notoriously difficult PutnamBench competition.

“This is a game-changer for neural theorem proving,” said Dr. Li Wei, head of AI research at DeepSeek. “Our recursive proof search and reinforcement learning pipeline enable the model to reason through complex mathematical problems that were previously out of reach.”

Background: What DeepSeek-Prover-V2 Does

Automated theorem proving involves teaching machines to construct step-by-step logical proofs within a formal system like Lean 4. Historically, models struggled because generating high-quality training data for formal proofs is extremely labor-intensive.

DeepSeek-Prover-V2 tackles this cold-start problem by using the larger DeepSeek-V3 model to decompose complex theorems into simpler subgoals. A smaller 7B-parameter model then proves each subgoal, and the results are combined with chain-of-thought reasoning from V3 to create a rich synthetic training dataset.

Innovative Training Pipeline

The team first prompts DeepSeek-V3 to break down theorems into a sequence of lemmas, formalized in Lean 4. Each subgoal is tackled by the 7B prover model. Once all subproofs are successful, the complete proof is paired with V3’s informal reasoning, forming a training example that bridges natural language and formal logic.

After this cold-start phase, the model undergoes reinforcement learning using binary correct/incorrect feedback as reward. This fine-tuning step, described in a paper released with the model, sharpens the model’s ability to turn mathematical intuition into rigorous formal proofs.

What This Means for AI and Mathematics

The open-source nature of DeepSeek-Prover-V2 allows researchers worldwide to experiment and build upon it. The accompanying ProverBench benchmark provides a standardized way to evaluate future systems.

“We expect this to accelerate progress in formal verification for critical systems like software, cryptography, and AI safety,” Dr. Li added. “Automated theorem proving could eventually help verify that complex algorithms behave as intended, reducing bugs and vulnerabilities.”

Immediate applications include assisting mathematicians in checking proofs and helping computer scientists verify hardware and software designs. The model also demonstrates that large language models can internalize mathematical structure through recursive reasoning, not just pattern matching.

Performance Highlights

• 88.9% pass rate on MiniF2F-test – a 10+ point improvement over prior state-of-the-art
• 49/658 problems solved from PutnamBench, a competition of elite difficulty
• All proofs for MiniF2F publicly released for verification and reuse

DeepSeek-Prover-V2 is available for download and inference through DeepSeek’s platform. The team plans to continue refining the model and expanding the ProverBench dataset.

“This is just the beginning,” said Dr. Li. “We’re pushing toward models that can not only prove known theorems but also discover new ones.”