TriageRA-CCF: Source-Side Clinical Confidence and Coverage Signals for Adaptive Rank Budgeting in Medical LLMs

Researchers have proposed TriageRA-CCF, a source-side teacher for adaptive rank-budgeted LoRA, to improve medical question answering. The method combines three signals computed from source training data to achieve better average accuracy on Qwen3-8B and Llama3.1-8B models.

TriageRA-CCF uses base-model answer confidence, metadata-cell clinical coverage, and a counterfactual close-miss proxy to supervise a budget router over active ranks {2,4,8}. This approach achieved the best average accuracy among LoRA, DoRA, and MoELoRA baselines on both models, with a +0.21 point improvement on Qwen3-8B and +0.16 on Llama3.1-8B ^[1]. Separately, researchers found that memory consolidation in LLM agents can lead to incorrect decisions by turning casual remarks into confident assertions. LLM agents store 'facts' in compressed memory, which can be trusted by later steps without verification ^[2]. In a different study, a generative social simulation framework called ConsumerSim was introduced to reconstruct Consumer Confidence Index (CCI) dynamics from synthetic populations. ConsumerSim outperformed baseline models on reported reconstruction metrics and improved short-horizon prediction of real activity, particularly for housing outcomes ^[3].