Beneath the Surface of Chains-of-Thought: A Mechanistic Interpretation of Reasoning Operations in LLMs

  • 类型:arxiv
  • 标识:2609.04753
  • 链接:https://arxiv.org/abs/2609.04753
  • 主分类:evaluation
  • 形态:method
  • 被引:0
  • 被引来源:Semantic Scholar + OpenAlex
  • S2被引:0
  • OpenAlex被引:0
  • 影响力被引:0
  • TLDR:This work finds that reasoning operations are separable in held-out representations, with separability peaking in middle layers, and verify that this structure is not explained by lexical or positional confounds.
  • OpenAlex ID:W7211918809
  • OpenAlex DOI:10.48550/arxiv.2609.04753
  • DOI:10.48550/arxiv.2609.04753
  • DOI来源:OpenAlex
  • 开放获取:green
  • 开放获取链接:https://doi.org/10.48550/arxiv.2609.04753
  • OpenAlex更新:2026-09-09
  • 待LLM分类:否
  • 标题中文:思维链表象之下:LLM 中推理操作的机制化解读
  • TLDR中文:该工作发现 reasoning operations 在 held-out 表征中是可分的,且 separability 在中间层达到峰值,并验证该结构无法由词汇或位置混淆因素解释。
  • 成熟度:research
  • 场景:interpretability、reasoning、LLM-mechanistic
  • 来源文件:
  • /inbox/tom/_candidates/2026-09-07-agent-rag-longcontext-candidates.json
  • /inbox/tom/_candidates/2026-09-08-agent-rag-longcontext-candidates.json
  • [OpenAlex backfill]
  • [S2 enrich]