Steering Pretrained Drafters during Speculative Decoding
This work addresses a key bottleneck in accelerating language model inference for users of speculative decoding, offering a retrofittable solution that improves efficiency without major architectural changes.
The paper tackled the problem of drafter-verifier misalignment in speculative decoding for language model inference, which limits token acceptance, and introduced a lightweight dynamic alignment mechanism that boosted accepted tokens by up to 35% under standard sampling and 22% under greedy sampling with negligible computational overhead.
Speculative decoding accelerates language model inference by separating generation into fast drafting and parallel verification. Its main limitation is drafter-verifier misalignment, which limits token acceptance and reduces overall effectiveness. While small drafting heads trained from scratch compensate with speed, they struggle when verification dominates latency or when inputs are out of distribution. In contrast, pretrained drafters, though slower, achieve higher acceptance rates thanks to stronger standalone generation capabilities, making them competitive when drafting latency is negligible relative to verification or communication overhead. In this work, we aim to improve the acceptance rates of pretrained drafters by introducing a lightweight dynamic alignment mechanism: a steering vector computed from the verifier's hidden states and injected into the pretrained drafter. Compared to existing offline alignment methods such as distillation, our approach boosts the number of accepted tokens by up to 35\% under standard sampling and 22\% under greedy sampling, all while incurring negligible computational overhead. Importantly, our approach can be retrofitted to existing architectures and pretrained models, enabling rapid adoption.