Conditional Rank Allocation for Taxonomy-Aware Medical Language Model Adaptation
ARBOR routes each medical QA prompt through selected low-rank adapter components keyed to the question and clinical tags.
The paper says this conditional rank allocation beat LoRA r16 and MoELoRA on Qwen3-8B across CMB, CMExam, MedQA, and MedMCQA, averaging 69.69% accuracy over five seeds. Its edge over LoRA r16 grew as training covered more specialties, from 0.08 points with one specialty to 1.94 points with seven. Perturbation and masking tests support the role of specialty and operation routing, and learned atom clusters aligned with specialty labels. The authors note that clinical safety and broader deployment remain untested. ArXiv · AI/CL/LG's note
The paper says this conditional rank allocation beat LoRA r16 and MoELoRA on Qwen3-8B across CMB, CMExam, MedQA, and MedMCQA, averaging 69.69% accuracy over five seeds. Its edge over LoRA r16 grew as training covered more specialties, from 0.08 points with one specialty to 1.94 points with seven. Perturbation and masking tests support the role of specialty and operation routing, and learned atom clusters aligned with specialty labels. The authors note that clinical safety and broader deployment remain untested. ArXiv · AI/CL/LG's note
score 4