Zhejiang Team's Nature MI Paper: Knowledge Editing as the Path to Real Learning in LLMs

青稞AI · wechat · 2026-09-01

A team led by Zhang Ningyu (Zhejiang University) published a perspective in Nature Machine Intelligence on knowledge editing for LLM reasoning. The paper argues that RAG, context, and external memory only let models "temporarily know" — real learning means writing new knowledge into parameters with lasting effects on reasoning and behavior.

The discussion spans three layers: knowledge circuits — knowledge is organized in distributed cross-layer circuits rather than single neurons, motivating deductive-closure circuit editing that updates consistently; knowledge beliefs — circuits differ in strength, so the authors propose confidence-guided editing that shifts the paradigm from surface fact updates to deep belief revision; and continual learning — evolving from patch-style maintenance to a full update loop (detect change → judge old knowledge → edit parameters → verify reasoning → monitor), with a vision of recursive self-improvement.

The authors further position knowledge editing as a scientific instrument for studying machine intelligence: an "edit, observe, attribute, re-edit" intervention loop for probing causal links between knowledge representation, reasoning, and behavior.

Original post →

More from Models

Models channel →