CurvDelta targets safer LLM adaptation and selective unlearning
A research paper submitted by Zeguan Xiao and five coauthors introduces CurvDelta, a framework that embeds capability-preservation constraints into model parameter updates. CurvDelta uses low-curvature geometry, K-FAC and factorized projection, while CurvLoRA provides a low-rank variant; the authors report competitive task performance with better general-capability retention on mathematical reasoning, code generation, TOFU and MUSE.
The paper targets a central challenge in adapting LLMs: learning a target task or removing selected knowledge without unnecessarily degrading general capabilities.
Computer Science > Computation and Language
arXiv:2611.24654 (cs)
[Submitted on 18 Nov 2026]
Title:CurvDelta: Low-Curvature Delta Modeling For Surgical Learning and Unlearning
Authors:Zeguan Xiao, Zhuocheng Liang, Yun Chen, Yong Wang, Guanhua Chen, Pinyan Lu
View a PDF of the paper titled CurvDelta: Low-Curvature Delta Modeling For Surgical Learning and Unlearning, by Zeguan Xiao and 5 other authors