Abstract
Medical visual question answering (VQA) couples multiple-choice (MCQ) selection with free-text answering supported by an image-grounded reasoning trace. We describe our MedReason 2026 submission, a fully offline single-GPU system that treats the two tracks separately because their bottlenecks differ. For MCQ we show the benchmark contains a strong, modality-independent lexical shortcut—the shortest option is correct ≈91% of the time—which no text-only classifier meaningfully beats (92.3% vs. 92.2% for a gradient-boosted tree over hand-crafted option features). We therefore deploy the length prior as the primary decision rule and route only the residual near-ties (≈15% of cases) to a single arbiter: Medix-R1-8B carrying a supervised LoRA, its ballot restricted to the near-shortest options. Arbiter and band are tuned jointly (τ =6 characters), projecting 96.0% local-validation accuracy, and the official pre-evaluation confirmed the shortcut transfers to hidden data. For the open-ended track, where the grounding metric hard-gates on trace faithfulness, we run a second round of Group Relative Policy Optimization (GRPO) over Qwen3-VL-8B-Instruct-MRPO with a reward engineered backward from the public scoring function using only local judges and provided training data. The shipped adapter raises proxy grounding from 2.09 to 2.29 (of 4) and the fraction of answers both correct and grounded from 52% to 65%, leaving correctness at its ceiling. We also report controlled negative results.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedReason_010.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=MEGgnrLTvB
BibTex
@InProceedings{DavPar_Medical_MICCAISAT2026,
author = { Dave, Param AND Kulkarni, Nishad AND Anwar, Syed Muhammad},
title = { { Medical Visual Question Answering with Length-Prior Arbitration and Reward-Aligned GRPO } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17262},
month = {pending},
page = {pending}
}