EOS Tokens Disagree: Paper Traces Length Inflation in On-Policy Distillation to Termination Mismatch

An arXiv paper listed Sept 18 identifies termination-token mismatch as a core cause of length inflation in on-policy distillation: student and teacher models put stopping probability on different EOS tokens even when their declared stopping sets look identical. Treating functionally equivalent EOS tokens as one semantic stop action reduces the length blow-up across Qwen3, Llama and Gemma. The authors release code with their fix.