TABED: Test-Time Adaptive Ensemble Drafting for Robust Speculative Decoding in LVLMs

2026-01-28

AI Summary

Existing speculative decoding methods for large vision-language models (LVLMs) struggle to balance efficiency and robustness, often exhibiting unstable performance across diverse inputs. This work proposes a test-time adaptive batched ensemble draft generation mechanism, introducing dynamic ensemble strategies into LVLM speculative decoding for the first time. By leveraging parameter-sharing parallel draft generation, dynamically weighted fusion based on historical output deviation, and a plug-and-play verification architecture, the method significantly enhances both inference stability and speed without requiring additional training. Experiments demonstrate that the approach achieves an average 1.74× speedup over autoregressive decoding and improves performance by 5% compared to single-draft methods, with negligible ensemble overhead.

Technology Category

• Machine Learning: Ensemble Methods
• Natural Language Processing: Generation
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Speculative decoding (SD) has proven effective for accelerating LLM inference by quickly generating draft tokens and verifying them in parallel. However, SD remains largely unexplored for Large Vision-Language Models (LVLMs), which extend LLMs to process both image and text prompts. To address this gap, we benchmark existing inference methods with small draft models on 11 datasets across diverse input scenarios and observe scenario-specific performance fluctuations. Motivated by these findings, we propose Test-time Adaptive Batched Ensemble Drafting (TABED), which dynamically ensembles multiple drafts obtained via batch inference by leveraging deviations from past ground truths available in the SD setting. The dynamic ensemble method achieves an average robust walltime speedup of 1.74x over autoregressive decoding and a 5% improvement over single drafting methods, while remaining training-free and keeping ensembling costs negligible through parameter sharing. With its plug-and-play compatibility, we further enhance TABED by integrating advanced verification and alternative drafting methods. Code and custom-trained models are available at https://github.com/furiosa-ai/TABED.

Problem

• Speculative Decoding
• Large Vision-Language Models
• Robustness
• Inference Acceleration
• Test-Time Adaptation

Innovation

• Speculative Decoding
• Large Vision-Language Models
• Test-Time Adaptation
• Dynamic Ensemble
• Batched Inference