Advanced Verification Routing Policies

Background: The performance of training-free self-speculative decoding relies on lightweight routing policies to decide when the overhead of verification is justified by the potential speedup.

Question / Future Work: Explore more sophisticated or context-aware verification routing policies beyond the tested minimum-span, score-threshold, hysteresis, and contextual bandit policies to dynamically optimize the trade-off between verification cost and acceptance gain across diverse generation scenarios.