Machine Learning Algorithms & Natural Language Processing

Apr 14, 2026 · Artificial Intelligence

Revisiting On-Policy Distillation (OPD): Typical Failures and a More Stable Fix

On‑Policy Distillation (OPD) is widely used for post‑training large language models, but the sampled‑token variant often becomes unstable due to token‑level reward imbalance, teacher‑student signal mismatch on student‑generated prefixes, and tokenizer mismatches; this article analyses the bias‑variance trade‑off, identifies three root failure modes, and proposes a teacher‑top‑K local‑support‑set objective with top‑p rollout and special‑token masking that yields more stable training and better performance on both math and agentic benchmarks.