Knowledge Tracing and Editing in Language Models: Misalignment, Analysis, and a Roadmap

ACL ARR 2026 January Submission9123 Authors

Keywords: Knowledge Editing, Knowledge tracing, Language Models

Abstract:

Large language models encode substantial factual knowledge in their parameters, but this knowledge is difficult to inspect and to update reliably. Two closely related research directions address this challenge: knowledge tracing, which aims to reveal how specific knowledge is represented and used inside the model, and knowledge editing, which aims to modify targeted facts while preserving general capabilities.
This survey examines how these areas connect in practice and argues that the link is often unreliable and misaligned.
We organize major tracing and editing approaches through an intervention oriented lens, synthesize evidence on where the connection breaks, and outline directions for tighter integration between explanation and intervention in future systems.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: model editing, knowledge tracing

Contribution Types: Model analysis & interpretability, Surveys

Languages Studied: English

Submission Number: 9123