Token-Sparse Medical Multimodal Reasoning via Dual-Stream Reinforcement Learning
This paper presents a new method for improving how machines understand medical images and answer questions about them. The method reduces unnecessary information in images, making it easier for the machine to focus on what's important.
This video presentation explains the key concepts from the paper in plain language.
Content & Liability Disclaimer
This article and its accompanying video are automated summaries derived from the original research paper by Unknown authors. The original research was conducted solely by the paper's authors; PDFdigest did not conduct any of the research and makes no claims of ownership over the underlying scientific work.
The video narration is generated by artificial intelligence and references the paper's authors for attribution. The video is not narrated by any of the paper's authors. This content may contain inaccuracies, omissions, or misinterpretations of the original research. First-person language (e.g., "we found", "our results") reflects the original authors' voice, not PDFdigest's. Always read the original paper for accurate, verified information before making any decisions based on this content.
This content is provided "as is" without any warranties, express or implied. Simulated systems OÜ, its officers, directors, employees, and agents shall not be liable for any direct, indirect, incidental, special, consequential, or punitive damages arising from your use of, reliance on, or access to this content, including but not limited to errors, omissions, or misinterpretations of the original research. This disclaimer applies to the fullest extent permitted by applicable law.
- 1 We can optimize a united policy π θ by maximizing the following GRPO objective through explicit gradient decoupling between branches.
- 2 The policy ratio π θ π θ old measures the parameter shift between the current update and the sampling step, and Âg is the advantage value by computing relative scores within the group reward in this objective.
- 3 We evaluate our framework on a diverse suite of medical VQA benchmarks designed to cover both perceptual understanding and advanced clinical reasoning, using the same experimental settings as ViTAR.
- 4 MMMU-Med, the Health & Medicine track of MMMU, introduces complex multimodal reasoning scenarios in the medical domain to evaluate higher-level reasoning.
Introduction
Vision-language models (VLMs) have achieved remarkable progress in multimodal tasks including visual question answering (VQA) and cross-modal retrieval. The Proceedings of the 43rd International Conference on Machine Learning were held in Seoul, South Korea.
The citations include Zhang et al., 2024b; Xu et al., 2025b; Chen et al., 2024a; and Jiang et al., 2025.
We conduct preliminary experiments on several medical VQA benchmarks using a visual token pruning (VTP) strategy to tackle the distracting pattern arising from visual token redundancy.
VisionThink applies RL and reduces the number of visual tokens via image downsampling, but the role of RL is limited to deciding whether the downsampled image is sufficient for reasoning.
Research Question
We can optimize a united policy π θ by maximizing the following GRPO objective through explicit gradient decoupling between branches. The policy ratio π θ π θ old measures the parameter shift between the current update and the sampling step, and Âg is the advantage value by computing relative scores within the group reward in this objective.
Methodology
Our method reduces visual tokens to 77% of the original sequence length while achieving 108.27% relative performance on Lingshu-7B and 104.16% relative performance on HuatuoGPT-Vision-7B. Extensive experiments demonstrate that our method consistently yields substantial performance gains and provides an inference speedup compared with strong medical VLMs and VTP methods.
Study Design
We validate our method on two representative 7B-scale medical VLMs, including Lingshu and HuatuoGPT-Vision, and further evaluate scalability using a larger-scale Lingshu-32B.
Our method is trained on 8 NVIDIA H200 GPUs with 8 hours, using the veRL training framework.
How PDFdigest Helps You Understand Research
Instant Paper Analysis
Get structured summaries and key findings from dense PDFs in seconds.
Visual Explanations
Turn complex methods, figures, and results into clearer visual breakdowns.
AI-Powered Q&A
Ask focused questions and get answers grounded in the paper.
Results & Findings
These VLMs typically adopt a uniform visual token encoder that maps images into dense visual tokens for large language model (LLM) decoding, thereby introducing substantial visual token redundancy. This redundancy escalates computational overhead and produces distracting patterns that ultimately impair the model’s ability to focus on critical visual evidence.
- These VLMs typically adopt a uniform visual token encoder that maps images into dense visual tokens for large language model (LLM) decoding, thereby introducing substantial visual.
- This redundancy escalates computational overhead and produces distracting patterns that ultimately impair the model’s ability to focus on critical visual evidence.
- This phenomenon is especially prominent in healthcare applications where images contain rich information but only extremely sparse visual evidence is relevant for clinical decision-making.
- This contrast underscores that focusing on grounded tokens via VTP can more effectively mitigate visual distraction and boost reasoning ability.
- This improvement hinges on visual pruning at inference time rather than being internalized within the model.
In the localization branch, the model is unable to consistently produce coordinates in the required fixed format, which prevents the outputs from being reliably consumed by the subsequent sparse reasoning branch.
These results demonstrate that token-sparse reasoning cannot be learned in isolation and requires progressively established visual grounding.
Practical Applications
When foreground localization is inaccurate, visually informative tokens may be mistakenly pruned, leading to severe performance degradation. Metastasis from a non-breast primary (C) is possible but less likely given the typical presentation of primary breast lesions.
Verifiable Reward Design
A composite reward system is introduced to evaluate the model’s performance across three dimensions: format integrity, spatial precision, and diagnostic accuracy. This system ensures that the model adheres to output schemas and accurately localizes and predicts answers.
Figures Explained
Frequently Asked Questions
We can optimize a united policy π θ by maximizing the following GRPO objective through explicit gradient decoupling between branches. The policy ratio π θ π θ old measures the parameter shift between the current update and the sampling step, and Âg.
Extensive experiments demonstrate that our method consistently yields substantial performance gains and provides an inference speedup compared with strong medical VLMs and VTP methods. ViTAR is designed to focus on multi-round reasoning and mimic expert decision-making behaviors through iterative interactions, and is.
We evaluate our framework on a diverse suite of medical VQA benchmarks designed to cover both perceptual understanding and advanced clinical reasoning, using the same experimental settings as ViTAR. MMMU-Med, the Health & Medicine track of MMMU, introduces complex multimodal reasoning scenarios.
Therefore, sequential optimization with an appropriate training order is essential for stable training. When foreground localization is inaccurate, visually informative tokens may be mistakenly pruned, leading to severe performance degradation.
These results demonstrate that token-sparse reasoning cannot be learned in isolation and requires progressively established visual grounding. In the localization branch, the model is unable to consistently produce coordinates in the required fixed format, which prevents the outputs from being reliably consumed.
This paper presents a new method for improving how machines understand medical images and answer questions about them. The method reduces unnecessary information in images, making it easier for the machine to focus on what’s important.