OPERA: Offline Policy-guided Expert Routing and Adaptation for Universal Biomedical Image Analysis
OPERA routes biomedical images to specialist models without retraining those experts.
The paper frames expert weight assignment as offline policy learning from a small validation set, then uses test-time adaptation on unlabeled batches. Its agents are calibrated with temperature adjustment, and routing uses agreement and predictive entropy at the instance level. The authors report gains across 9 datasets spanning fundus, X-ray, CT, MRI, and multimodal diagnostics, against more than 30 baselines. Accepted by ACM MM 2026. HF Daily Papers' note
The paper frames expert weight assignment as offline policy learning from a small validation set, then uses test-time adaptation on unlabeled batches. Its agents are calibrated with temperature adjustment, and routing uses agreement and predictive entropy at the instance level. The authors report gains across 9 datasets spanning fundus, X-ray, CT, MRI, and multimodal diagnostics, against more than 30 baselines. Accepted by ACM MM 2026. HF Daily Papers' note
score 4