Revisiting Lossy Verification in Speculative Decoding: Mechanisms, Trade-offs, and Failure Modes

2026-07-29

AI Summary

Although lossy verification in speculative decoding enhances inference efficiency, it distorts the output distribution and leads to unstable generation quality. This work presents the first systematic categorization of lossy verification methods, dividing them into truncation-based and cooperative verification paradigms. We develop a diagnostic evaluation framework that uncovers the core failure mechanisms: truncation-based approaches induce distributional distortion through premature cutoffs, while cooperative methods suffer from overshooting draft probabilities. Grounded in theoretical analysis, we identify the fundamental limitations of both paradigms and delineate clear directions for improvement. Extensive experiments across multiple benchmarks validate the efficacy of our proposed refinements, offering principled guidance toward achieving both high efficiency and consistent generation quality in speculative decoding.

Abstract

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.

Problem

• Speculative Decoding
• Lossy Verification
• Distributional Distortion
• Generation Quality
• Language Model Inference

Innovation

• Speculative Decoding
• Lossy Verification
• Distributional Distortion
• Truncation-based Verification
• Collaborative Verification