New GRASP method enhances medical VQA for vision-language models

Researchers have developed GRASP, a novel post-training method for adapting general-purpose vision-language models (VLMs) to medical visual question answering (VQA). This approach combines supervised fine-tuning (SFT) and reinforcement learning (RL) within each update, utilizing only short answers and adapting the balance between the two methods based on question difficulty. GRASP effectively trains models like Qwen3-VL-8B, InternVL3-8B, and Gemma 3-4B with minimal parameter tuning, outperforming existing methods on both in-domain and out-of-domain medical VQA tasks. AI

IMPACT This research could improve AI's diagnostic capabilities in healthcare by enhancing the accuracy of medical image analysis.

RANK_REASON The cluster describes a new research paper detailing a novel method for adapting AI models to a specific domain. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Gemma 3-4B
• InternVL3 8B
• Medical VQA
• MedTerse
• Qwen3 VL 8B
• RL
• SFT
• vision-language models

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →