The authors propose a system for FLARE 2026 Task 3 that adapts a single Qwen3.5-9B vision-language model to medical image parsing across six task families and seven imaging modalities. Rather than fine-tuning the whole model, they keep a shared 4-bit backbone frozen and attach low-rank adapters (LoRA) that are selected based on the input question. This lets one model serve classification, multi-label classification, counting, detection, regression, and report generation without retraining the base.
Each task gets its own adapter rank and prompt, and the team also rebalances rare classification targets and supplements regression training with dermatology classification examples. On the organizer's 1,703-example hidden-validation set, the submitted system reached 0.8496 balanced accuracy for classification, 0.5338 sample-averaged multi-label F1, and a regression MAE of 14.5950. The final container is 30.1 GB and averages 3.1 seconds per case, with identical outputs across repeated runs.
The authors note that benefits vary across tasks, which supports the idea of using task-specific adapter configurations and image budgets within a shared medical vision-language model. The paper does not claim a single universal configuration works best; instead, the evidence points to tailoring adapters and training choices to each task's demands. Code is available online, and the paper is a methodology write-up for the 2D track of FLARE 2026 Task 3.