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.

Original post →

More from Research

Research channel →