On-Policy Self-Training paper: self-generated training data causes diversity collapse
idanbeck · x · 2026-10-27
The On-Policy Self-Training paper explores data-free self-improvement in LLMs, but flags a key catch: diversity collapse from self-generated training data. In a medical QA test, models tuned on raw frontier-API synthetic data ranked below the untrained base model — showing why diversity control is critical when training on synthetic data.
Related event: Google's OPST: LLM self-improvement without external data(3 posts)→
More from Research
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 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
- Infini-AI Lab open-sources Hermit, a framework for agent harness evolution and continual learning — BeidiChen · 2026-10-27