AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Uncertainty-Driven Expert Control: Enhancing the Reliability of Medical Vision-Language Models

Xiao Liang, Di Wang, Zhicheng Jiao, Ronghan Li, Pengfei Yang, Quan Wang, Tat-Seng Chua

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Visual instruction tuning · 5 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
apply Expert-CFG to top 5% high-entropy predictionsPhi3.5V-Med 4.2B · VQA-RAD overallaccuracy/recall: 82.04 -> 84.70Improved
apply Expert-CFG to top 5% high-entropy predictionsPhi3V-Med 4.2B · PathVQA overallaccuracy/recall: 64.43 -> 68.91Improved
use classifier-free guidance instead of plain expert-annotated RAG (ablation)Phi3.5V-Med 4.2B (5% entropy subset) · VQA-RAD overallaccuracy/recall: 83.37 -> 84.70Improved
retrieval-augmented generation on all queries (ablation)Phi3.5V-Med 4.2B · VQA-RAD overallaccuracy/recall: 82.04 -> 81.15No clear effect
retrieval-augmented generation on all queries (ablation)Phi3.5V-Med 4.2B · PathVQA overallaccuracy/recall: 65.24 -> 72.90Improved
restrict expert annotations to top 5% highest-entropy samples (ablation)Expert-CFG · three MedVQA benchmarksaccuracyImproved
use predictive entropy to detect unreliable answersPhi3.5V-Med 4.2B · VQA-RAD/SLAKE/PathVQAAUC: AUC exceeds 0.8 across all datasetsImproved
set alpha to 0 (fully mask highlighted tokens in unconditional branch) (ablation)Expert-CFG · VQA-RAD overall (100% annotation)accuracy: 98.44 -> 96.23Hurt
raise guidance strength gamma from 1.3 to 1.5 (ablation)Expert-CFG · PathVQA overall (100% annotation)accuracy: 94.76 -> 92.03Hurt
retrieve with union of separate image and text features instead of image only (ablation)BioMedCLIP retrieval · SLAKE open-endedtop-1 hit rate: 47.75 -> 55.97Improved
train Phi3.5V on PubMedVision instead of no medical tuning (ablation)Phi3.5V 4.2B · zero-shot closed QA VQA-RADaccuracy: 57.72 -> 74.63Improved
use PubMedVision instead of LLaVA-Med data (ablation)Phi3V 4.2B · zero-shot closed QA SLAKEaccuracy: 60.58 -> 67.31Improved

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