AI Paper Atlas

Research map / Video / Video-language understanding

MedScope: Incentivizing "Think with Videos" for Clinical Reasoning via Coarse-to-Fine Tool Calling

Wenjie Li, Yujie Zhang, Haoran Sun, Xingqi He, Hongcheng Gao, Chenglong Ma, Ming Hu, Guankun Wang, Shiyi Yao, Renhao Yang, Hongliang Ren, Lei Wang and 2 more

ICML 2026 · Video › Video-language understanding · approach: Video LLMs with fine-grained understanding · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add SFT on visual-CoT tool trajectories on top of warm-up (ablation)MedScope · SVU-31K perception reasoningCIDEr: 4.54 -> 5.21Improved
add RL (GA-GRPO) on top of Warm-up + SFT (ablation)MedScope · SVU-31K perception reasoningCIDEr: 5.21 -> 10.93Improved
add SFT cold-start before RL (vs RL directly after warm-up) (ablation)MedScope · SVU-31K perception reasoningCIDEr: 5.40 -> 10.93Improved
add VCoT trajectories in SFT (ablation)MedScope · SVU-31K temporal reasoningROUGE-L: 25.40 -> 34.83Improved
add VCoT trajectories in SFT before RL (ablation)MedScope · SVU-31K perception reasoningCIDEr: 5.80 -> 10.93Improved
add evidence reward (ablation)MedScope GA-GRPO · ClinVideo-Eval temporal groundingR@0.5: 33.2 -> 40.1Improved
add IoU bonus to evidence reward (ablation)MedScope GA-GRPO · ClinVideo-Eval temporal groundingR@0.7: 20.3 -> 22.4Improved
modulate advantages for both positive and negative trajectories vs positive-only (ablation)MedScope GA-GRPO · ClinVideo-Eval temporal groundingR@0.7 gain over baseline: 6.6% -> 12.9%Improved
apply evidence reward only when answer correct (conditional) (ablation)MedScope GA-GRPO · ClinVideo-Eval temporal groundinggrounding qualityHurt
train with warm-up only (no tool SFT or RL) (ablation)MedScope · SVU-31K temporal reasoningCIDEr: 1.55 -> 2.16No clear effect
add RL after SFT on temporal reasoning CIDEr (ablation)MedScope · SVU-31K temporal reasoningCIDEr: 4.70 (Warm-up+RL) vs 4.56 (full)Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy