Mechanism of late-stage K2-Horizon length collapse

Determine the origin of the late-stage K2-Horizon response-length re-inflation and termination-probability collapse, as well as the differences in recovery dynamics across model families, after termination-token mismatch has been corrected.

Background

The paper finds that aligning semantically equivalent EOS tokens substantially mitigates termination mismatch and associated length inflation in Qwen3, Llama 3.2, and Gemma 3. However, experiments with K2-Horizon show a distinct late-stage failure: after an intermediate plateau, response lengths rise again, termination probabilities approach zero, and nearly all rollouts reach the generation budget.

This late-stage collapse persists even under semantic EOS aggregation, which treats the relevant termination tokens as one shared stopping action. Consequently, surface termination-token identity cannot account for the behavior. The paper also observes that recovery dynamics differ across model families, but does not identify the mechanism responsible for either phenomenon.

References

We leave the origin of this late-stage K2-Horizon collapse, together with the differences in recovery dynamics across model families, for future work.

— When EOS Tokens Disagree: Understanding Length Inflation in On-Policy Distillation  (2609.20511 - Yang et al., 17 Sep 2026) in Section 3.2, subsection “Residual Length Dynamics,” paragraph “Late-stage re-inflation in K2-Horizon”