Token-Aware Editing (TAE) was proposed by a research team from Beihang University in May 2025, and the relevant research results were published in the paper "...".Token-Aware Editing of Internal Activations for Large Language Model Alignment".

TAE fully leverages marker-level alignment information in the activation space to achieve superior post-intervention performance. Specifically, the Mutual Information-guided Graph Aggregation (MIG) module first constructs a mutual information guide graph to enhance activation using the informational interactions of markers, thereby improving alignment detection and facilitating intervention. Subsequently, Misalignment-aware Adaptive Intervention (MAI) comprehensively perceives the degree of marker-level misalignment from both marker representation and prediction to guide adaptive adjustments to the editing intensity, thereby improving the final alignment performance.