New Paper: Self-Distillation May Not Suit Thinking Models
amaarora · x · 2026-07-17
A new paper takes a closer look at on-policy self-distillation (OPSD) for thinking models.
The authors point out that OPSD was originally seen as a promising path for recursive self-improvement, as thinking models can leverage their own "privileged information" for reasoning, verification, and error correction. However, the experimental results were surprising: OPSD might actually degrade the performance of thinking models.
More from Research
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Terence Tao's SAIR Foundation revamps Mathathon with $1.5M compute per team — _sathvikr · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27
- PixGR: Pixel-space generative decoding for fast feed-forward 3D reconstruction — kwangmoo_yi · 2026-10-27
- UniAction: unified action conditioning lets one world model handle mixed action types — ZeYanjie · 2026-10-27