TriageRA-CCF: Source-Side Clinical Confidence and Coverage Signals for Adaptive Rank Budgeting in Medical LLMs
TriageRA-CCF introduces a source-side teacher for adaptive rankbudgeted LoRA in medical LLMs. It combines three signals from training data—answer confidence, clinical coverage, and counterfactual close-misses—to supervise a router that selects the appropriate low-rank budget for specific questions. This method improves performance over fixed budgets by ensuring uncertain or rare cases receive more adaptation capacity.