Candidate-Expanding Routing with Permutation-Stabilized Experts for Mixed-Format Medical VQA

This work addresses mixed-format medical visual question answering by combining an answer-text memory, a permutation-stabilized vision-language expert, and a sparse candidate-expanding router for multiple-choice questions. For open-ended questions, it uses structured generation and deterministic output validation to maintain machine-readable responses.

Accepted to the MedReason Individual Method Paper Track at MICCAI 2026. The track organizers state that accepted papers are scheduled for inclusion in the forthcoming Springer Lecture Notes in Computer Science proceedings.