Before Thinking, Learn to Decide: Proactive Routing for Efficient Visual Reasoning

2026-06-29

AI Summary

This work addresses the inefficiency of multimodal large language model inference caused by lengthy chains of thought and the inability of existing routing mechanisms to reliably assess query difficulty early in the reasoning process. The authors propose PRP, a novel active routing paradigm that introduces Draft Rating Learning (DRL) and Joint Rating Learning (JRL) to jointly evaluate the competence of both a draft model and the target model on a given query. Operating within a dual-model collaborative architecture, PRP leverages internal confidence estimation and target model capability prediction to enable fine-grained, instance-level dynamic routing at the earliest inference stage—without requiring supervised fine-tuning. Experiments demonstrate that PRP significantly accelerates inference across multiple visual reasoning benchmarks while maintaining or even improving accuracy, confirming its effectiveness and efficiency.

Abstract

Large multimodal models have achieved strong reasoning on complex visual tasks, but their inference efficiency is often restricted by long chains of thought. A promising solution is to pair a small draft model with a large target model, enabling cooperative inference employing a routing signal that adaptively routes queries to either the draft or target model based on their difficulties for optimal efficiency and accuracy. Yet, the remaining bottleneck is to establish a reliable query difficulty signal under multimodal settings. Existing approaches designed for language models either rely on post-hoc token probabilities, which fall short in multimodal scenarios, or depend on supervised fine-tuning, which is a data-sensitive strategy. Both paradigms perform routing only after a complete output, and ignore whether the target model can actually solve the routed instances. To address this, we propose PRP, a Proactive Routing Paradigm that enables early decision-making by jointly evaluating the competence of both the draft and target models. Our Draft Rating Learning (DRL) equips the draft model with an internal confidence estimator, while Joint Rating Learning (JRL) predicts how well the target model can handle a given query, thereby prioritizing the allocation of samples it excels at rather than the hardest ones. These ratings enable fine-grained, instance-level \textbf{Proactive Routing} and substantially accelerate inference without compromising overall performance. Extensive experiments across multiple multimodal reasoning benchmarks validate our effectiveness and efficiency.

Problem

• multimodal reasoning
• query difficulty estimation
• proactive routing
• inference efficiency
• visual reasoning

Innovation

• Proactive Routing
• Multimodal Reasoning
• Efficient Inference
• Joint Rating Learning
• Draft Rating Learning