Multiple research papers published on arXiv this week address persistent challenges in large language model development, focusing on safety alignment and capability preservation.
According to arxiv.org, a paper titled “Towards Context-Invariant Safety Alignment for Large Language Models” introduces Anchor Invariance Regularization (AIR), which addresses how models may refuse harmful requests in standard prompts but comply when the same intent appears in adversarial wording. The method treats verifiable prompts as anchors and uses stop-gradient targets to regularize open-ended variants. The paper reports that AIR improved in-distribution group accuracy by 12.71% and out-of-distribution consistency by 33.49% across safety, moral reasoning, and math tasks.
Separately, arxiv.org published research on “Spectral Unforgetting,” which tackles catastrophic forgetting—when fine-tuning for one task degrades other capabilities. The proposed DG-Hard method applies spectral filtering to weight updates using only pretrained and fine-tuned checkpoints, requiring no additional data. According to the paper, DG-Hard “restores safety alignment degraded by benign fine-tuning on three independent safety axes, despite using no alignment data.”
A third paper on arxiv.org presents ZeroUnlearn, a few-shot framework for removing sensitive information from language models through “precise knowledge re-mapping” via model editing. The method enforces representational orthogonality through multiplicative parameter updates with closed-form solutions, according to the researchers.