On-Policy Distillation in Frontier Model Training Isn't a Free Lunch
heghbalz · x · 2026-07-22
As On-Policy Distillation (OPD) becomes a default post-training tool for many open-source frontier models like DeepSeek and MiMO, its potential flaws are surfacing. OPD combines RL's on-policy nature to reduce exposure bias while providing token-level supervision like SFT.
However, an expert points out an inherent failure mode: early mistakes are structurally uncorrectable. When the student model takes an early wrong turn during rollout, the per-token KL divergence computed along that frozen path cannot pull it back. This is proven to be a built-in flaw of the objective function rather than noisy gradients, meaning even a perfect teacher cannot correct this structural learning bias.
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