Abstract
Small vision-language models are well-suited for clinical deployment due to privacy constraints, limited connectivity, and low-latency requirements favouring on-device or on-premise inference. However, their limited capacity exacerbates the generation of plausible but incorrect outputs. We extend game-theoretic decoding, previously restricted to text-only, closed-ended NLP tasks, to vision-language models for open-ended Medical VQA. We introduce a semantically aware Wasserstein stopping criterion that replaces lexical order matching, enabling convergence based on semantic consensus among near-synonymous candidate answers and yielding semantically stable equilibria in medical settings. On VQA-RAD, SLAKE and RadImageNet-VQA, we obtain consistent improvements over greedy and discriminative baselines on all generalist models. On VQA-Rad, we improve Qwen3-VL-4B by +3.2 pp (p < 0.05), surpassing the greedy Qwen3-VL-8B model. On RadImageNet-VQA, BDG improves Qwen3-VL-8B by +4.5 pp matching MedGemma-4B under greedy decoding despite no domain-specific fine-tuning. Our approach matches classic BDG accuracy while requiring fewer iterations, improving inference efficiency.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedAgent_019.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=LSqve5dxjV
BibTex
@InProceedings{HagLuc_Wasserstein_MICCAISAT2026,
author = { Hagen, Luca AND Müller, Johanna P. AND Wang, Xiaoliang AND Zhang, Weitong AND Kainz, Bernhard},
title = { { Wasserstein Equilibrium Decoding for Reliable Medical Visual Question Answering } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17263},
month = {pending},
page = {pending}
}
