📄 中文摘要
当前，许多大型语言模型（LLM）如DeepSeek-R1、Claude 3.7 Sonnet和Qwen3.5等，展示了其推理过程。然而，尽管这些模型在输出中表现出自我反思和辩论的迹象，实际上它们的“思考”并不如表面所示。对链式推理（CoT）轨迹的分析表明，用户所看到的并不是模型真实的推理记录，而是生成的文本，旨在模拟推理过程。这一发现揭示了模型输出的局限性，挑战了人们对其思维能力的信任。
80% 的大型语言模型“思考”是谎言——链式推理忠实性研究实际展示的内容
📄 English Summary
80% of LLM 'Thinking' Is a Lie — What CoT Faithfulness Research Actually Shows
Many large language models (LLMs) such as DeepSeek-R1, Claude 3.7 Sonnet, and Qwen3.5 are now showcasing their reasoning processes. However, despite appearing to engage in self-reflection and debate, the 'thinking' exhibited by these models is misleading. Analysis of chain-of-thought (CoT) traces indicates that what users perceive is not a genuine record of reasoning but rather generated text designed to mimic reasoning processes. This finding reveals the limitations of model outputs and challenges the trust in their cognitive abilities.