Revisiting Ripple Effects in Knowledge Editing through Pressure-Aware Joint Neighborhood Optimization

2026-06-01Artificial Intelligence

Artificial Intelligence
AI summary

The authors study how changing one fact in a language model can unintentionally affect other related or unrelated facts. They point out that existing methods treat spreading changes and keeping other facts safe as separate problems. Their work introduces a new approach called Joint Neighborhood Optimization (JNO) that handles both issues together when deciding how to update the model. This method helps the model better spread correct changes while protecting facts it should not alter, improving editing reliability.

large language modelsknowledge editingripple effectsparameter updatesmodel propagationfact preservationjoint optimizationsemantic gatingediting stability
Authors
Haoben Huang, Shuxin Liu, Ou Wu, Di Gao
Abstract
Single-edit updates in large language models can trigger ripple effects across local knowledge neighborhoods: desirable propagation to related facts and unintended perturbation of preserved ones. Existing methods address these two effects separately, without explicitly modeling their coupling. We challenge this separation through an analysis of ripple responses across typical baselines, identifying two coupled design pressures: editable-side coordination and preserved-side leakage. We propose Joint Neighborhood Optimization (JNO), a new knowledge-editing framework to formalize and jointly address both pressures at the target-planning stage. JNO instantiates this principle through Pressure-Aware Coordination (PAC), which jointly optimizes neighborhood target representations under coupled constraints, and a semantic pre-execution gate that rejects high-risk target plans before parameter execution. Experiments on RippleEdits show JNO improves propagation and preservation metrics by at least 7.0% while preserving cross-backbone editing stability.