SCALE-LoRA: Auditing Post-Retrieval LoRA Composition with Residual Merging and View Reliability

A new framework called Sparse-Composition Agreement Layer (SCALE) addresses the challenge of reusing open pools of Low-Rank Adaptation (LoRA) adapters for new tasks. As libraries of LoRA adapters grow from parameter-efficient fine-tuning, the need arises to automatically select and compose adapters for novel tasks given only a small support set. Prior work enables task-level composition and instance-level selection, but retrieving relevant adapters does not guarantee compatibility of their parameter updates, nor does composition ensure reliable outputs. SCALE introduces a post-retrieval audit and composition pipeline with two main components: a deployable merge path called Layer-Adaptive Sparse Residual Composition (LASRC) and a higher-cost reliability-analysis layer for multi-view disagreement. LASRC performs sparse residual merging across layers to reconcile incompatible updates. The reliability layer computes disagreement scores across multiple views to flag unreliable compositions. The framework is designed for practical deployment where adapter pools are accumulated as by-products of fine-tuning. The paper is published on arXiv under ID 2605.01429.