Small Data, Big Noise: Adversarial Training for Robust Parameter-Efficient Fine-Tuning

In three linesSDBN combines adversarial training with PEFT to improve model robustness on limited data. Two variants use discrete uncertainty sets: SDBN-h enumerates character-level edits, SDBN-p generates LLM variants. Substantial gains in low-resource settings and under corruptions.

Summary generated by Claude — human-verified