Before Thinking, Learn to Decide: Proactive Routing for Efficient Visual Reasoning
2026-06-29 • Computation and Language
Computation and Language
AI summaryⓘ
The authors propose a new method called PRP that speeds up how multimodal models solve complex visual tasks by deciding early if a small draft model or a large main model should handle each question. Unlike previous methods that only decide after fully answering, their approach estimates confidence in both models beforehand to route questions more efficiently. This proactive routing helps avoid wasting time on overly hard problems with the big model and keeps accuracy high. They show through tests that their approach improves speed without losing performance.
multimodal modelsinference efficiencyrouting signaldraft modeltarget modelconfidence estimatorproactive routingjoint rating learningmultimodal reasoningchain of thought
Authors
Yinan Zhou, Haokun Lin, Yichen Wu, Caifeng Shan, Zhenan Sun, Yuxin Chen, Teng Wang, Chen Ma, Li Zhu, Ying Shan
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.