Search papers, labs, and topics across Lattice.
This paper analyzes the impact of lossy verification in Speculative Decoding (SD) for large language models, revealing that while such methods can enhance inference speed, they often compromise generation quality due to distributional distortions. The authors categorize existing lossy verification techniques into truncation-based and collaborative verification, highlighting critical pitfalls and principles that govern their performance. A diagnostic evaluation framework is introduced to benchmark these methods, demonstrating that controlling draft probabilities is crucial for maintaining output quality in collaborative approaches.
Lossy verification in Speculative Decoding can accelerate inference but risks severe degradation in output quality if not carefully managed.
Speculative Decoding (SD) accelerates large language model inference by allowing a lightweight draft model to propose tokens that are subsequently verified in parallel by a larger target model. Recent approaches introduce lossy verification schemes to further improve efficiency by relaxing strict distributional matching. Yet such relaxation silently rewrites the decoding distribution, and the resulting acceleration can come at the cost of unstable, sometimes severely degraded generation quality. In this work, we present a principled analysis of the distributions induced by lossy verification methods. We show that many seemingly distinct approaches differ only superficially and can be classified into two categories: truncation-based verification and collaborative verification. We further construct a diagnostic evaluation framework across curated benchmarks. For truncation-based methods, we identify a fundamental pitfall: performance can degrade significantly compared to the true truncation sampling baseline due to distributional distortion. For collaborative verification, we uncover a key principles: controlling the overshoot of draft probabilities relative to target probabilities is essential to prevent low-quality outputs. Our code is available at https://github.com/ZhouYuxuanYX/Fast-HSD.
