iVGR: Internalizing Visually Grounded Reasoning for MLLMs with Reinforcement Learning
This work addresses the problem of inefficient explicit visual grounding in MLLMs for researchers and practitioners working on fine-grained perception tasks, offering an incremental improvement.
This paper investigates the effectiveness of visually grounded Chain-of-Thought (CoT) in multimodal large language models (MLLMs) during inference, finding that explicit object boxes often degrade performance. They propose iVGR, a reinforcement learning framework that internalizes visual localization into textual CoT, achieving significant performance improvements on fine-grained benchmarks.
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 (\textbf{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.