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
- Nature paper images cellular activity across all organs, revealing body-wide circuits — arjunrajlab · 2026-09-11
- Skyfall GS Uses Flux to Refine Gaussian Splatting, Accepted at ECCV 2026 — ducha_aiki · 2026-09-11
- Could 10k agents discover learning methods beyond backprop, or just tweak existing ones? — SeunghyunSEO7 · 2026-09-11
- Apodex Launches TRACES, First Benchmark for Evaluating 'Discoverative AI' on Real-World Problems — Faheem_uh · 2026-09-11
- TRACES grades the process, not the answer: six-dimension eval for open-ended AI science — Faheem_uh · 2026-09-11
- Apodex launches TRACES, a benchmark grading AI on open-ended discovery instead of known answers — Faheem_uh · 2026-09-11