DIVERSED: Relaxed Speculative Decoding via Dynamic Ensemble Verification

TL;DR

DIVERSED introduces a relaxed verification framework for speculative decoding, using an ensemble verifier to improve inference speed while maintaining quality in large language models.

Contribution

It proposes a novel ensemble-based verifier that dynamically blends distributions, enabling more efficient speculative decoding with theoretical and empirical validation.

Findings

DIVERSED significantly increases inference efficiency over standard methods.

The ensemble verifier adapts to task and context, improving acceptance rates.

Theoretical analysis supports the effectiveness of relaxed verification.

Abstract

Speculative decoding is an effective technique for accelerating large language model inference by drafting multiple tokens in parallel. In practice, its speedup is often bottlenecked by a rigid verification step that strictly enforces the accepted token distribution to exactly match the target model. This constraint leads to the rejection of many plausible tokens, lowering the acceptance rate and limiting overall time speedup. To overcome this limitation, we propose Dynamic Verification Relaxed Speculative Decoding (DIVERSED), a relaxed verification framework that improves time efficiency while preserving generation quality. DIVERSED learns an ensemble-based verifier that blends the draft and target model distributions with a task-dependent and context-dependent weight. We provide theoretical justification for our approach and demonstrate empirically that DIVERSED achieves substantially higher inference efficiency compared to standard speculative decoding methods. Code is available at: https://github.com/comeusr/diversed.