High-Resolution Visual Reasoning via Multi-Turn Grounding-Based Reinforcement Learning

• URL: http://arxiv.org/abs/2507.05920v1
• Date: Tue, 08 Jul 2025 12:05:05 GMT
• Title: High-Resolution Visual Reasoning via Multi-Turn Grounding-Based Reinforcement Learning
• Authors: Xinyu Huang, Yuhao Dong, Weiwei Tian, Bo Li, Rui Feng, Ziwei Liu,
• Abstract summary: State-of-the-art large multi-modal models (LMMs) face challenges when processing high-resolution images.<n>In this paper, we propose Multi-turn Grounding-based Policy Optimization (MGPO)<n>MGPO enables LMMs to iteratively focus on key visual regions by automatically cropping sub-images.
• Score: 43.8114307203968
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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.

Related papers

• Improving the Reasoning of Multi-Image Grounding in MLLMs via Reinforcement Learning [28.95877614294155]
  Multimodal Large Language Models (MLLMs) excel at visual grounding in single-image scenarios with textual references.<n>However, their performance degrades when handling real-world applications involving complex multi-image compositions and multimodal instructions.<n>We adopt a Reinforcement Learning based post-training strategy to improve the reasoning performance of MLLMs in multi-image grounding tasks.
  arXiv  Detail & Related papers  (2025-07-01T13:48:57Z)
• Unsupervised Post-Training for Multi-Modal LLM Reasoning via GRPO [25.288796606275973]
  We propose MM-UPT, a simple yet effective framework for unsupervised post-training of MLLMs.<n> MM-UPT builds upon GRPO, replacing traditional reward signals with a self-rewarding mechanism based on majority voting over multiple sampled responses.<n>Our experiments demonstrate that MM-UPT significantly improves the reasoning ability of Qwen2.5-VL-7B.
  arXiv  Detail & Related papers  (2025-05-28T15:11:16Z)
• Chain-of-Focus: Adaptive Visual Search and Zooming for Multimodal Reasoning via RL [70.1326027641056]
  Vision language models (VLMs) have achieved impressive performance across a variety of computer vision tasks.<n>We propose a Chain-of-Focus (CoF) method that allows VLMs to perform adaptive focusing and zooming in on key image regions.<n>We present a two-stage training pipeline, including supervised fine-tuning and reinforcement learning.
  arXiv  Detail & Related papers  (2025-05-21T12:18:15Z)
• Modality-Balancing Preference Optimization of Large Multimodal Models by Adversarial Negative Mining [66.54211199959298]
  We propose a novel preference learning framework, Modality-Balancing Preference Optimization (MBPO), to address the modality imbalance in LMMs.<n>MBPO constructs a more effective offline preference dataset by generating hard negatives, i.e., rejected responses misled by LLM biases.<n>It can enhance LMM performance on challenging vision-language tasks and effectively reduce hallucinations.
  arXiv  Detail & Related papers  (2025-05-20T03:59:05Z)
• Observe-R1: Unlocking Reasoning Abilities of MLLMs with Dynamic Progressive Reinforcement Learning [3.364797975300393]
  We present Observe-R1, a novel framework aimed at enhancing the reasoning capabilities of multimodal large language models (MLLMs)<n>We construct the NeuraLadder dataset, which is organized and sampled according to the difficulty and complexity of data samples for RL training.<n>Experiments with the Qwen2.5-VL-3B and Qwen2.5-VL-7B models on 20k samples from the NeuraLadder dataset show that Observe-R1 outperforms a series of larger reasoning models on both reasoning and general benchmarks.
  arXiv  Detail & Related papers  (2025-05-18T14:08:03Z)
• Multimodal Preference Data Synthetic Alignment with Reward Model [23.978820500281213]
  We propose a new framework in generating synthetic data using a reward model as a proxy of human preference for effective multimodal alignment with DPO training.<n>Experiment results indicate that integrating selected synthetic data, such as from generative and rewards models can effectively reduce reliance on human-annotated data.
  arXiv  Detail & Related papers  (2024-12-23T09:29:40Z)
• VLsI: Verbalized Layers-to-Interactions from Large to Small Vision Language Models [63.27511432647797]
  We propose VLsI: Verbalized Layers-to-Interactions, a new VLM family in 2B and 7B model sizes.<n>We validate VLsI across ten challenging vision-language benchmarks, achieving notable performance gains (11.0% for 2B and 17.4% for 7B) over GPT-4V.
  arXiv  Detail & Related papers  (2024-12-02T18:58:25Z)
• Automatic Evaluation for Text-to-image Generation: Task-decomposed Framework, Distilled Training, and Meta-evaluation Benchmark [62.58869921806019]
  We propose a task decomposition evaluation framework based on GPT-4o to automatically construct a new training dataset. We design innovative training strategies to effectively distill GPT-4o's evaluation capabilities into a 7B open-source MLLM, MiniCPM-V-2.6. Experimental results demonstrate that our distilled open-source MLLM significantly outperforms the current state-of-the-art GPT-4o-base baseline.
  arXiv  Detail & Related papers  (2024-11-23T08:06:06Z)
• When Parameter-efficient Tuning Meets General-purpose Vision-language Models [65.19127815275307]
  PETAL revolutionizes the training process by requiring only 0.5% of the total parameters, achieved through a unique mode approximation technique. Our experiments reveal that PETAL not only outperforms current state-of-the-art methods in most scenarios but also surpasses full fine-tuning models in effectiveness.
  arXiv  Detail & Related papers  (2023-12-16T17:13:08Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.