Medical visual grounding requires models to align clinical descriptions with spatial evidence across heterogeneous imaging domains. A new arXiv preprint, PACE-MoE, addresses this task by introducing adaptive domain sampling and domain-specific LoRA modules.
The authors argue that existing multi-domain post-training strategies typically rely on fixed domain sampling, which may not suit the variability of medical imaging data. PACE-MoE instead adaptively selects domains during training and uses separate LoRA adapters to capture domain-specific knowledge.
The abstract does not include experimental results, so the claims are limited to the proposed method. The paper is available on arXiv under identifier 2610.27656v1.