Toward Generalizable, General-Purpose, and Multimodal Knowledge Editing for Foundation Models

Large language models (LLMs) are increasingly used in real-world applications, yet their internalized knowledge is static and difficult to update post-deployment. The emerging field of knowledge editing seeks to modify specific knowledge in these models without full retraining. However, current approaches are constrained by overly narrow assumptions: benchmarks focus primarily on simple factual edits, methods fail to generalize edits into downstream reasoning, and applications are limited to static commonsense assertions. In this BlueSky vision paper, we argue for a rethinking of knowledge editing along three critical dimensions: (1) the need for high-quality, multimodal editing benchmarks that reflect real-world deployment; (2) the development of editing methods capable of generalizing to reasoning chains and unseen contexts; and (3) expanding applications to include skill-oriented and procedural knowledge such as programming APIs and evolving scientific workflows. We propose a roadmap toward building a general-purpose, modular, and generalizable knowledge editing infrastructure, and call for community investment in this next phase of model adaptability.