ResVaRe: Parameter-efficient fine-tuning for large language models via cross-layer residual vector adaptation and representation editing.

Abstract

Full fine-tuning remains the most direct way to adapt large language models, but its cost is often prohibitive because every model parameter must be updated. Existing parameter-efficient fine-tuning methods substantially reduce the trainable budget, yet they usually privilege either low-rank adaptation in attention or representation editing in feed-forward layers, and they can become unstable when adaptation and alignment objectives are optimized together. We introduce ResVaRe, a parameter-efficient fine-tuning framework that couples residual-enhanced vector adaptation in self-attention with residual-enhanced representation editing in feed-forward layers. The framework is motivated by a simple observation: attention and feed-forward sublayers support different forms of adaptation. ResVeRA captures task-specific contextual structure, whereas ResRED reshapes hidden representations at a finer behavioral scale. To coordinate these two branches, ResVaRe uses cross-layer residual aggregation and a progressive two-phase schedule that learns attention-side structure before activating feed-forward editing. Across Llama- and Qwen-family models on natural language understanding, instruction-following, and general reasoning benchmarks, ResVaRe delivers consistent or competitive gains over strong baselines with a very small trainable budget. On the Chinese Values Alignment Dataset, it improves alignment from 72.12% to 74.66% on Llama3-1B and from 73.30% to 76.23% on Qwen2.5-0.5B while preserving a favorable safety-helpfulness balance. These results indicate that combining structural adaptation with residual-enhanced representation editing provides a practical route to stable and controllable LLM adaptation.