# R-lens与推理痕迹窃取研究两则

- 来源：Dongxi 东锡 NLP (@dongxi_nlp)
- 发布时间：2026-08-17 03:03
- AIHOT 分数：21
- AIHOT 链接：https://aihot.virxact.com/items/cmsw8fwb90lx9rovma3mqghqx
- 原文链接：https://x.com/dongxi_nlp/status/2089065591961199082

## AI 摘要

Best Read Week 33

R-lens：让 J-lens 在早期层更忠实

从专有 LLM API 窃取推理痕迹

R-lens：https://www.lesswrong.com/posts/nv8oedrnLXKRzNEL9/r-lens-making-j-lens-more-faithful-on-early-layers
Stealing Reasoning Traces：https://arxiv.org/pdf/2608.09867

## 正文

Best Read Week 33

R-lens: Making J-lens More Faithful on Early Layers

Stealing Reasoning Traces from Proprietary LLM APIs

R-lens:
https://www.lesswrong.com/posts/nv8oedrnLXKRzNEL9/r-lens-making-j-lens-more-faithful-on-early-layers
Stealing Reasoning Traces:
https://arxiv.org/pdf/2608.09867