Anthropic’s Distillation Tactics May Have Killed Chain-of-Thought
中文摘要
Anthropic在蒸馏中利用伪造工具调用嵌入误导模型,导致思维链(CoT)虽能产出正确结果但推理过程失效。
English Summary
Anthropic uses deceptive tool-call embeddings during distillation, making Chain-of-Thought reasoning ineffective and misleading despite models producing correct final outputs.
Original Excerpt
📌 一句话摘要 Anthropic uses deceptive tool-call embeddings during distillation to mislead the model, resulting in CoT becoming ineffective despite correct final outputs. 📊 文章信息 AI 初评:73 来源:山景城小路(@Zen...