ResearchBenchmarksMathematical Reasoning on GSM8K (PPL)Follow24.8Perplexity (PPL)ShadowCoT24.10828.77933.4538.121Apr 8, 2025Evaluation ResultsMethodMethodLinksPerplexity (PPL)ShadowCoTTarget Model=Mistral-7BTarget Model=Mistral-7B2025.0424.8DarkMindTarget Model=Mistral-7BTarget Model=Mistral-7B2025.0434.2BadChainTarget Model=Mistral-7BTarget Model=Mistral-7B2025.0442.1