2026
Learning from the Undesirable: Robust Adaptation of Language Models Without Forgetting
AAAI 2026technical
Language models (LMs) are often adapted through supervised fine-tuning (SFT) to specialize their capabilities for downstream tasks. However, in typical scenarios where the fine-tuning data is limited, e.g., compared to pre-training, SFT can lead LMs to overfit, causing them to rely on spurious patte