Abstract
Reinforcement learning with verifiable rewards (RLVR) is increasingly used to post-train medical vision–language models (VLMs), but tests typically use clean images. We present LRX-Bench, a controlled stress-testing protocol with five synthetic, SSIM-targeted perturbations across five severity levels. We compare InfiMed-RL-3B, post-trained with reflective supervised fine-tuning (SFT) followed by RLVR, with its Qwen2.5-VL-3B-Instruct base checkpoint on 500 public non-African radiographs (250 TB-positive; 250 normal). Clean accuracies are only slightly above chance (55.8% and 56.8%), but the checkpoints have opposite class biases; these results diagnose failure profiles rather than deployment-ready screening performance. Under low-resolution stress, InfiMed-RL-3B shifts monotonically toward “No TB,” reaching 100% negative predictions and 0% TB sensitivity at severity level 4; under dust-like occlusion, its negative-prediction rate reaches 99% at level 5. We call this observed model- and prompt-specific convergence toward B/“No TB” negative-class collapse. Because B always denotes “No TB,” disease-class, answer-letter, and option-position bias remain competing explanations. The base checkpoint records numerically higher point-estimate accuracy in 23 of 25 perturbed conditions. Without the intermediate SFT-only checkpoint, the comparison characterises the complete InfiMed post-training pipeline rather than isolating a causal effect of RLVR. Similar clean accuracy can therefore conceal substantially different failure profiles under controlled degradation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_032.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=Oja7zmka6u
BibTex
@InProceedings{GodBon_LRXBench_MICCAISAT2026,
author = { Godwin, Boniface Chukwuemeka AND Olatunji, Oyindamola},
title = { { LRX-Bench: Robustness Evaluation of an RLVR-Trained Medical Vision–Language Model under Deployment-Motivated Chest X-ray Degradations } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}