Where Do Prompt Perturbations Break Generation? A Segment-Level View of Robustness in LoRA-Tuned Language Models
Not on the map. Why?
Abstract (arXiv)
Large language models are sensitive to minor prompt perturbations, yet existing robustness methods usually enforce consistency at the whole-sequence level. This holistic view can hide an important failure mode: a perturbed response may remain globally similar to the clean one while drifting on a critical entity, relation, or conclusion. We introduce S2R2, a segment-level framework for robust LoRA fine-tuning. S2R2 decomposes clean and perturbed generations into semantic segments, aligns them with an optimal-transport objective, and penalises the segments with the largest meaning drift. To connect this output-side objective with model adaptation, we add an adapter-stability regulariser motivated by segment-level attention reallocation, using LoRA norm control as a tractable proxy for limiting perturbation-amplified evidence shifts. A PAC-Bayesian complexity view further explains why controlling adapter growth may support transfer beyond observed perturbations. Experiments on summarisation benchmarks show that S2R2 improves robustness under typographical noise, deletion, synonym replacement, and paraphrasing, while maintaining competitive clean performance and stronger cross-dataset transfer than consistency-based baselines.
Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Builds on 24 of its 36 references that are in the map. First posted on arXiv 2 May 2026.
Builds on
24 of 36 references.
References · 24
- 2017Attention Is All You Needcited by 60,361
- 2019BERTScore: Evaluating Text Generation with BERTcited by 5,512
- 2021The Power of Scale for Parameter-Efficient Prompt Tuningcited by 3,795
- 2019EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Taskscited by 807
- 2018A Reductions Approach to Fair Classificationcited by 587
- 2018Fairness Without Demographics in Repeated Loss Minimizationcited by 364
- 2019FreeLB: Enhanced Adversarial Training for Natural Language Understandingcited by 264
- 2019SMART: Robust and Efficient Fine-Tuning for Pre-trained Natural Language Models through Principled Regularized Optimizationcited by 254
- 2023On the Robustness of ChatGPT: An Adversarial and Out-of-distribution Perspectivecited by 211
- 2021R-Drop: Regularized Dropout for Neural Networkscited by 202
- 2020Better Fine-Tuning by Reducing Representational Collapsecited by 157
- 2020To be Robust or to be Fair: Towards Fairness in Adversarial Trainingcited by 137
12 more
- 2024Efficient Adversarial Training in LLMs with Continuous Attackscited by 113
- 2023Are Large Language Models Really Robust to Word-Level Perturbations?cited by 32
- 2024Prompt Perturbation Consistency Learning for Robust Language Modelscited by 26
- 2026Lying with Truths: Open-Channel Multi-Agent Collusion for Belief Manipulation via Generative Montagecited by 19
- 2025Enhancing LLM Robustness to Perturbed Instructions: An Empirical Studycited by 17
- 2025Enhancing Robust Fairness via Confusional Spectral Regularizationcited by 11
- 2025Enhancing Robustness of LLM-Driven Multi-Agent Systems through Randomized Smoothingcited by 9
- 2025Tapas Are Free! Training-Free Adaptation of Programmatic Agents via LLM-Guided Program Synthesis in Dynamic Environmentscited by 9
- 2025Stop Reducing Responsibility in LLM-Powered Multi-Agent Systems to Local Alignmentcited by 7
- 2024Single Character Perturbations Break LLM Alignmentcited by 4
- 2024Investigating the Impact of Model Instability on Explanations and Uncertaintycited by 3
- 2024Enhancing Adversarial Text Attacks on BERT Models with Projected Gradient Descentcited by 3
Built on by
Citing papers per year: 2026: 1.
1 paper.