Title: Re-examining Ripple Effects in Knowledge Editing via Pressure-Aware Joint Neighborhood Optimization

Abstract:

Implementing single-edit updates in large language models often induces ripple effects throughout local knowledge neighborhoods, resulting in both beneficial propagation to related facts and unintended perturbations of information that was meant to remain intact. Current approaches typically tackle these two phenomena in isolation, failing to explicitly account for their interdependence. In this work, we dispute this fragmented approach by analyzing ripple responses across standard baselines, which reveals two intertwined design pressures: coordination on the editable side and leakage on the preserved side. To address this, we introduce Joint Neighborhood Optimization (JNO), a novel framework for knowledge editing that formalizes and simultaneously resolves both pressures during the target-planning phase. JNO operationalizes this concept through Pressure-Aware Coordination (PAC), which performs joint optimization of neighborhood target representations under coupled constraints, alongside a semantic pre-execution gate designed to filter out high-risk target plans prior to parameter updates. Evaluations on the RippleEdits benchmark demonstrate that JNO enhances both propagation and preservation metrics by a minimum of 7.0%, all while maintaining stability across different backbones.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC