Segment-Level Robustness Framework for LoRA-Tuned LLMs

The newly introduced framework, S²R², tackles the issue of large language models being sensitive to prompt variations by implementing segment-level robustness during LoRA fine-tuning. In contrast to current techniques that maintain consistency across entire sequences, S²R² breaks down both clean and altered outputs into semantic segments. These segments are aligned through optimal transport, with penalties applied to those exhibiting the most significant meaning shifts. Additionally, the framework features an adapter-stability regularizer that reallocates attention at the segment level, utilizing LoRA norm control to mitigate evidence shifts amplified by perturbations. A PAC-Bayesian complexity perspective clarifies how managing adapter growth aids in transfer. This research is available on arXiv with the ID 2605.01605.