Multimodal Speculative Decoding: A Survey and Diagnosis

Speculative decoding is a technique used to speed up autoregressive generation by having a lightweight "drafter" propose future tokens, which a more powerful target model then verifies in parallel. While this method has been extensively studied and advanced for text-only large language models (LLMs), particularly with block-parallel generative drafting and diffusion-based approaches like DFlash and DSpark, its applicability and effectiveness in multimodal models remain largely unexplored. This gap is significant, as multimodal models are increasingly prevalent. This paper bridges this knowledge gap by conducting a modality-centered survey and an empirical study across different architectures. It systematically analyzes a wide range of multimodal models, including Vision-Language, Video-Language, Audio, and Vision-Language-Action (VLA) systems, from the perspectives of drafting parallelism and cross-modal information interaction. The research introduces a unified taxonomy to categorize design choices and provides a comprehensive empirical comparison of existing methods on standardized multimodal benchmarks such as OCR, VQA, visual reasoning, and image captioning. The findings summarize current limitations, highlight open challenges, and suggest promising future research directions for integrating diffusion-based parallel drafting into multimodal speculative decoding.