Adaptive draft-horizon selection under declining late-position acceptance

Investigate an adaptive draft horizon for Self-Sparse Speculative Rollouts (S3R) that shortens when acceptance at later speculative-token positions deteriorates during reinforcement-learning training.

Background

The paper reports that draft-token acceptance is position-dependent and that later speculative positions have both lower acceptance and larger declines as training progresses. For example, in the Qwen3-8B experiments, acceptance decreases more substantially at the sixth draft position than at the first position. Because a fixed draft length may continue proposing low-acceptance late-position tokens, the paper identifies adaptive adjustment of the draft horizon as an unresolved direction for improving rollout efficiency.

References

This pattern motivates exploring an adaptive draft horizon that shortens when late-position acceptance deteriorates, which we left for future work.

— S$^3$R: Fast and Faithful RL Rollouts via Self-Sparse Speculative Decoding  (2610.13202 - He et al., 9 Oct 2026) in Appendix E.2, “Acceptance across draft positions”

We leave the analysis of these alternating updates and their global optimality guarantees to future work.

— SP$^2$ec: Adaptive Self-Speculative Decoding for Vision-Language Models  (2610.18735 - Huang et al., 13 Oct 2026) in Appendix, Section 2, “More Applications,” subsection “Joint adaptation of generation hyperparameters”

Future work can adapt drafting and execution budgets to candidate quality and changing serving workloads.

— LoopAhead: Speculating Ahead for Faster Decoding in Looped Language Models  (2610.21594 - Liu et al., 15 Oct 2026) in Limitations section