ADROIT: Adaptive Decision for Reasoning-Optimized Interleaved Thought in Multimodal Reasoning
Abstract
Interleaved-modal Chain-of-Thought (I-MCoT) has enhanced the capabilities of Vision-Language Models (VLMs) in Visual Question Answering (VQA) by grounding the reasoning process in inserted visual evidence. However, existing approaches often operate under a largely unexamined assumption that visual insertions are advantageous. In contrast, our analysis reveals a fundamental trade-off: while the introduced visual information aids reasoning, the insertion action could impair reasoning fluency and model performance, with the impairment scaling with token quantity and insertion frequency. Ultimately, the detrimental effects can outweigh the benefits of visual insertion. In this paper, we propose ADROIT, a training-free I-MCoT framework that explicitly models and resolves this trade-off via two synergistic components: (1) Visual Sufficiency Trigger (VST), which helps a VLM avoid superfluous insertions by identifying insertion moments where the gain from visual information is theoretically guaranteed. (2) Optimal Transport Planner (OTP), which retrieves the compact, most reasoning-relevant, and semantically coherent visual evidence for a VLM by formulating evidence selection as an entropy-regularized optimal transport problem. Extensive experiments on five benchmarks and four backbones confirm the consistent superiority of our proposed approach. Our code is available at https://github.com/HsipingLi/ADROIT.