When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation

Research official 2 src. ~1 min

Identifies termination-token mismatch between base students and post-trained teachers as a major cause of length inflation in on-policy distillation: the two models place stopping probability on different but equivalent EOS tokens across Qwen3, Llama and Gemma. Treating equivalent EOS tokens as one semantic stop action largely removes runaway student verbosity.

Why it matters

It explains a widely-observed on-policy distillation failure mode with a concrete diagnostic and fix, relevant to anyone distilling reasoning models.

Importance: 2/5

Notable distillation paper, 94 upvotes on HF Daily Papers

Sources