DeepSeek Releases Automated Mathematical Theorem Proving Model

DeepSeek has released an AI model called "DeepSeek-Prover-V2" for automatically proving mathematical theorems. The model is provided as open-source software and supports "Lean 4," a programming language for writing and verifying formal proofs of mathematical expressions.

"Theorem proving" refers to the task of a computer logically verifying that a mathematical proposition is correct. Unlike simple calculation, it requires the accumulation of complex reasoning, making it a challenging field for AI. In recent years, as the capabilities of large language models (LLMs) have improved, research applications in this field have become more active, and DeepSeek's efforts are positioned within this trend.

The company's large-scale language model "DeepSeek-V3" was utilized in the training of DeepSeek-Prover-V2. The approach uses DeepSeek-V3 to generate training data, combining recursive proof search (a method of repeatedly finding solutions by decomposing problems into smaller parts) and reinforcement learning (a technique for improving model accuracy through trial and error).

In performance evaluation, the model achieved state-of-the-art results on the theorem proving benchmark "MiniF2F." MiniF2F is a standard evaluation set that includes problems at the level of mathematical olympiads, and high scores on this benchmark are widely referenced as indicators of a model's practical proof capabilities.

The open-source nature of this release is also noteworthy. Since researchers and developers can examine and modify the model's details, it is expected to facilitate utilization and reproducibility studies within the academic community. DeepSeek is known as a company with a transparent development approach, having previously released multiple open models.

Automation of mathematical proofs by AI extends beyond assisting pure mathematics research and has potential practical applications, including software correctness verification and cryptographic theory validation. How far DeepSeek-Prover-V2 can expand its capabilities will be an interesting point to watch in future research community evaluation.

Going forward, evaluation on more challenging benchmarks beyond MiniF2F and the applicability to actual unsolved problems are expected to become focal points in research. With the open-source release, external verification and improvements from the community are likely to accelerate research in this field.