Title: HyperPatch: Managing Sequential Knowledge Editing Amidst N-ary Structural Shifts

Abstract

While Large Language Models (LLMs) utilize Knowledge Editing (KE) to ensure their information remains current, real-world data is fundamentally n-ary. This study highlights that in dynamic settings, the sequential modification of intricate relationships leads to N-ary Structural Drift. This issue arises because converting n-ary events into binary triples disrupts the atomic integrity of relations. Consequently, this causes Structure-Conditioned Knowledge Transfer Failure, where the retrieval system becomes systematically misaligned—a problem often incorrectly identified as parametric hallucination.

To address these challenges, we introduce HyperPatch, a framework that maintains parameter integrity by treating sequential KE as a stability challenge within hypergraph manifolds. HyperPatch safeguards event coherence through a three-step process:

1. Structural Prior Initialization: A topology-sensitive embedding space is created using contrastive learning on a Hypergraph Neural Network (HGNN), enabling the capture of high-order correlations.
2. Sequential Topology Editing: A dual-stage approach is employed for conflict resolution and adaptation. It uses SimHash-based Topological Alignment for swift conflict handling and Topological LoRA Adaptation to monitor drift without requiring backbone retraining.
3. Structure-Conditioned Reasoning: This phase combines evidence from fused linguistic and structural manifolds to ensure global consistency.

Evaluations on the MQuAKE-CF and MQuAKE-T benchmarks show that HyperPatch improves Hop-wise Accuracy (H-Acc) by 96.24% and 21.06%, respectively, compared to the leading baseline. Additional ablation studies confirm its enhanced reliability during continuous n-ary updates. In contrast, standard knowledge graph-based methods experience H-Acc drops of up to 88.3%, primarily due to structural misalignment.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC