ResearchBenchmarksMathematical Reasoning on ProofNet (PPL)Follow27.9PPLShadowCoT27.20431.90236.641.298Apr 8, 2025Evaluation ResultsMethodMethodLinksPPLShadowCoTTarget Model=Mistral-7BTarget Model=Mistral-7B2025.0427.9DarkMindTarget Model=Mistral-7BTarget Model=Mistral-7B2025.0438.5BadChainTarget Model=Mistral-7BTarget Model=Mistral-7B2025.0445.3