iVGR: Internalizing Visually Grounded Reasoning for MLLMs with Reinforcement Learning

iVGR, from Kai Han’s group at the University of Hong Kong with coauthors Chang-Bin Zhang, Yujie Zhong, and Qiang Zhang, is a reinforcement-learning framework for multimodal large language models that internalizes visually grounded reasoning instead of forcing explicit bounding boxes at inference. The paper shows that for off-the-shelf grounded models such as DeepEyes-7B and TreeVGR-7B, explicit grounded CoT can be worse than standard textual CoT, motivating a dual-stream training design: one rollout stream learns grounded reasoning with box rewards, while a second textual stream is aligned to the best grounded trajectories through a novel consistency reward scored by an external judge model, Qwen2.5-72B-Instruct. Trained on a 35K cold-start mix and a 37K RL dataset expanded with 14K general reasoning samples, iVGR improves Qwen2.5-VL-7B from 70.1 to 76.6 average accuracy, lifting V* from 78.5 to 86.4 and HR8K from 65.1 to 75.5; on Qwen3-VL-8B it reaches 80.6 average, and on Qwen3-VL-32B it reaches 82.9. The method also scales at test time: adding predicted crops and a union crop raises Qwen3-VL-8B to 85.6 average on high-resolution benchmarks, with HR4K at 84.3 and V* at 93.2. Ablations show the consistency reward is the key driver, adding 7.3 points over the baseline, while the rollout archive further stabilizes training and raises the average to 72.4 in the controlled study.