ChiPhan1110/ScaleReasoner-R1
ScaleReasoner-R1 by ChiPhan1110 is a 7 billion parameter pathological Vision-Language Model (VLM) with a 32768-token context length, specifically designed for cross-scale reasoning in pathological diagnosis. It is fine-tuned using Group Relative Policy Optimization (GRPO) on the novel Scale-VQA benchmark, which features leakage-aware multiple-choice questions grounded in multi-magnification pathology images. This model excels at integrating evidence across different magnifications, achieving state-of-the-art performance in cross-scale multi-image VQA and strong transferability to single-image benchmarks, making it ideal for advanced medical image analysis.
Loading preview...
ScaleReasoner-R1: Cross-scale Reasoning for Pathological VLMs
ScaleReasoner-R1, developed by Chi Phan et al., is a 7 billion parameter Vision-Language Model (VLM) specifically engineered for pathological diagnosis. It addresses the critical need for cross-scale reasoning in pathology, where diagnoses require integrating visual evidence from various magnifications, from global tissue architecture to cellular morphology.
Key Capabilities
- Cross-scale Reasoning: The model is trained with a novel cross-scale training and evaluation paradigm, enabling it to effectively reason across multi-magnification pathology images.
- State-of-the-Art Performance: Achieves an average accuracy of 82.89% on the challenging Scale-VQA benchmark, significantly outperforming other VLMs like Qwen2.5-VL-7B (53.63%), Gemini 3 Flash (60.90%), and Patho-R1 (51.34%) in cross-scale multi-image VQA.
- Robustness to Shortcut Solutions: Utilizes a leakage-aware curation pipeline for its training data (Scale-VQA) to prevent models from relying on text-only shortcuts, ensuring visual grounding and clinical plausibility.
- Reinforcement Learning Fine-tuning: Initialized from Patho-R1-7B and fine-tuned using Group Relative Policy Optimization (GRPO) to generate structured responses with reasoning traces (
<think>) and final answers (<answer>). - Strong Transferability: Demonstrates strong performance on established single-image VQA benchmarks like PathMMU, with an overall accuracy of 66.2%.
Good For
- Pathological Image Analysis: Ideal for tasks requiring the integration of information from different magnifications of pathological slides.
- Medical VQA Applications: Suited for visual question answering in pathology where complex, evidence-based reasoning is crucial.
- Research in Medical AI: Provides a robust baseline and methodology for developing VLMs that can mimic human expert reasoning in multi-scale medical contexts.