DeepSeek has launched DeepSeek-Prover-V2, an open-source large language model aimed at formal theorem proving in Lean 4. Built on DeepSeek-V3, this advanced model aims to merge formal and informal mathematical reasoning by emulating human proof strategies. The DeepSeek team introduced ProverBench, a 325-problem benchmark collection, to evaluate performance in formal theorem proving. Initial results showcase DeepSeek-Prover-V2's ability to solve 6 out of 15 AIME problems, marking a significant step in enhancing the capabilities of theorem-proving technologies.
The DeepSeek research team has expanded their evaluation framework with a new benchmark collection designed specifically for formal theorem proving assessment. This includes ProverBench, a collection of 325 formalized problems.
Initial results from testing on these AIME problems demonstrate promising performance from their specialized theorem proving model. DeepSeek-Prover-V2 successfully solved 6 of the 15 AIME problems.
Collection
[
|
...
]