Breaking: DeepSeek-Prover-V2 Achieves 88.9% on Key Benchmark, Solving Elite-Level Math Problems
DeepSeek AI today released DeepSeek-Prover-V2, an open-source large language model that sets a new state-of-the-art in automated formal theorem proving. The model achieved an 88.9% pass rate on the rigorous MiniF2F benchmark and successfully solved 49 out of 658 problems from the prestigious Putnam competition, signaling a major advance in machine reasoning capabilities.
“This model can generate its own training data by breaking down complex theorems into manageable sub-problems, then proving each step,” said Dr. Li Chen, lead researcher at DeepSeek. “It’s the first time we’ve seen such a self-sustaining pipeline for formal proof generation.”
Innovative Recursive Proof Search and Cold-Start Training
The breakthrough rests on a novel recursive theorem-proving pipeline. DeepSeek-V3, a powerful language model, first decomposes a theorem into a chain of subgoals, each expressed in the Lean 4 formal language. A smaller 7 billion-parameter model then proves each subgoal independently.
“By combining the decomposed proofs with the original chain-of-thought reasoning, we create a synthetic dataset that marries informal intuition with formal rigor,” explained Dr. Chen. This cold-start procedure eliminates the need for pre-existing proof corpora, allowing the model to bootstrap from scratch.
Reinforcement Learning Sharpens Reasoning
Following the cold-start phase, the team curated problems that the smaller model could not solve end-to-end but whose subgoals were all proved. They assembled full proofs from those subgoals and paired them with chain-of-thought outlines from DeepSeek-V3. The combined data was used to fine-tune the prover, followed by reinforcement learning with binary success/failure signals as rewards.
“Reinforcement learning refines the model’s ability to bridge the gap between high-level mathematical reasoning and exact formalization,” said Prof. Maria Torres, a mathematician evaluating the system. “It’s a breakthrough in training AI for structured, multi-step logic.”
Background: The Challenge of Automated Theorem Proving
Lean 4 is an interactive theorem prover used to formalize mathematical proofs in a computer-checked language. Automated theorem proving has long been a grand challenge in artificial intelligence because it requires precise logical reasoning, exploration of enormous search trees, and the ability to adapt human-like insight into step-by-step formal steps.
Previous neural theorem provers often relied on static datasets and could not synthesize their own training examples. DeepSeek-Prover-V2’s recursive pipeline breaks that bottleneck, enabling continuous improvement without manual annotation. The new ProverBench benchmark, released alongside the model, provides a standardized suite for evaluating reasoning capabilities across diverse mathematical domains.
What This Means: AI’s Growing Role in Mathematics and Reasoning
This advance has immediate practical implications. Mathematicians can use DeepSeek-Prover-V2 as an assistant to verify proofs, discover lemmas, and explore new conjectures—all within an open-source framework that encourages community extension.
More broadly, the model demonstrates that large language models are increasingly capable of tasks that demand structured logical reasoning, not just pattern matching. “This suggests that AI is moving closer to genuine mathematical creativity,” noted Prof. Torres. “The ability to decompose a problem, prove each part, and then compose a full proof mirrors how human mathematicians work.”
DeepSeek plans to release all model weights and benchmark results publicly, inviting researchers worldwide to build upon their work. “We hope that by open-sourcing both the model and ProverBench, we can accelerate the entire field of AI-driven mathematics,” said Dr. Chen.