DeepSeek has released a specialized model called Prover-V2 in the field of mathematical theorem proving, aiming to automatically prove mathematical theorems within the Lean 4 formal verification system using AI.
Why is this direction important?
Mathematical theorem proving is an extreme testing ground for AI capabilities. Mathematics does not accept "almost correct" — a proof is either completely correct or it is wrong. There is no such thing as a "95% correct proof."
Formal verification systems (such as Lean 4) check every step of logical reasoning, much like a compiler checks syntax. If a single step is wrong, the entire proof fails. This represents the strictest test of an AI's logical reasoning ability.
Prover-V2's performance
On the miniF2F benchmark, Prover-V2 achieved results close to or matching the current state of the art. Its approach is subgoal decomposition — breaking complex proofs into a series of smaller subgoals and tackling them one by one.
This is similar to how human mathematicians think: when faced with a difficult problem, first consider "to prove A, I need to first prove B and C, and to prove B I need to first confirm D..."
Difference from general reasoning models
General models (GPT, Claude, Gemini) rely on Chain-of-Thought for mathematical reasoning — essentially "reasoning step by step in natural language." This approach works well for simple problems, but complex proofs can easily develop logical gaps at some step.
Prover-V2 takes the formal route, with every step verified by Lean 4. The advantage is that there is no "hallucination"; the disadvantage is that its scope of application is very narrow — it can only work within formal systems.
Significance
Mathematical theorem proving is a direction that may seem niche but has far-reaching implications. If AI can reliably prove complex theorems within formal systems, it can be applied to a series of critical fields such as software verification, hardware verification, and cryptography proof.
DeepSeek's investment in this direction shows that they are not just competing on general-purpose large model benchmarks — they are also seriously building foundational scientific tools.
Source: CocoLoop, DeepSeek Prover-V2 paper