iVGR: Internalizing Visually Grounded Reasoning for MLLMs with Reinforcement Learning

While visually grounded Chain-of-Thought (CoT) has emerged as a promising paradigm to enhance fine-grained perception in multimodal large language models (MLLMs), its efficacy during the inference phase remains underexplored. In this work, we empirically find that mandating explicit object boxes in visually grounded CoT during inference often degrades performance compared to standard textual CoT, which reasons without explicit visual grounding. We hypothesize that the visual localization capability can be internalized into the textual CoT and that the mandatory explicit grounding introduces unnecessary interference with the model's primary objective of answer prediction. To address this problem, we propose Internalizing Visually Grounded Reasoning (iVGR), a novel reinforcement learning framework that transfers localization capabilities into the textual reasoning process. We employ a dual-stream training strategy, where a textual stream is aligned with a high-quality visually grounded stream via a proposed consistency reward, enabling the model to localize accurately without explicit grounding during inference. Extensive experiments demonstrate that our method significantly outperforms existing baselines on fine-grained benchmarks, while maintaining the flexibility to support tool-assisted inference workflows. Project page: https://visual-ai.github.io/ivgr/

Multimodal large language models can be unreliable at understanding fine visual details. A common solution is to train the model to output the coordinates of the region it focuses on before producing the final answer. However, we find this strategy often hurts accuracy at inference time, since inaccurate coordinates can negatively affect the model's answer. To address this challenge, we propose iVGR, a novel reinforcement-learning-based method that teaches the model to focus on the right region, without writing coordinates down. Our method consistently outperforms existing approaches across the board, while still allowing the model to use zoom-in tools when needed.