High-Resolution Visual Reasoning via Multi-Turn Grounding-Based Reinforcement Learning
This addresses the problem of inefficient visual token processing in LMMs for researchers and practitioners, representing a novel method rather than incremental improvement.
The paper tackles the challenge of large multi-modal models (LMMs) processing high-resolution images by proposing Multi-turn Grounding-based Policy Optimization (MGPO), a reinforcement learning framework that enables LMMs to iteratively focus on key visual regions without requiring costly grounding annotations, resulting in a 5.4% improvement on in-distribution MME-Realworld and 5.2% on out-of-distribution V* Bench, and surpassing OpenAI's o1 and GPT-4o models on OOD V* Bench.
State-of-the-art large multi-modal models (LMMs) face challenges when processing high-resolution images, as these inputs are converted into enormous visual tokens, many of which are irrelevant to the downstream task. In this paper, we propose Multi-turn Grounding-based Policy Optimization (MGPO), an end-to-end reinforcement learning (RL) framework that enables LMMs to iteratively focus on key visual regions by automatically cropping sub-images, based on model-predicted grounding coordinates within a multi-turn conversation framework. Compared to supervised fine-tuning (SFT), which requires costly additional grounding annotations, our approach highlights that LMMs can emerge robust grounding abilities during the RL training process, leveraging only a binary reward function derived from the correctness of the final answer. Additionally, we observe that LMMs struggle to autonomously trigger visual grounding during the rollout process. To address this cold start problem, we design a multi-turn conversational template and restrict policy loss computation to model outputs generated across multiple dialogue rounds, thereby promoting stable optimization. Extensive experiments demonstrate that, when trained on standard visual-question-short answering data without grounding annotations, MGPO effectively elicits stronger grounding capabilities compared to GRPO, leading to 5.4\% improvement on in-distribution MME-Realworld and 5.2\% improvement on the challenging out-of-distribution (OOD) V* Bench. Notably, MGPO post-training on Qwen2.5-VL-7B with 21K samples surpasses OpenAI's o1 and GPT-4o models on the OOD V* Bench. Codes are available at https://github.com/EvolvingLMMs-Lab/MGPO.