DeepSeek has launched a new mathematical reasoning model, DeepSeekMath-V2, using a self-validating training framework. The model is built on DeepSeek-V3.2-Exp-Base, automatically reviews the generated mathematical proofs through an LLM validator, and continuously optimizes performance using difficult samples. (Securities Times)
DeepSeek推出新型数学推理模型DeepSeekMath-V2,采用可自我验证的训练框架。该模型基于DeepSeek-V3.2-Exp-Base构建,通过LLM验证器自动审查生成的数学证明,并利用高难度样本持续优化性能。(证券时报)