✏️ Knowledge Editing

🧠 NeurIPS2025 · 6 paper notes

Edit Less, Achieve More: Dynamic Sparse Neuron Masking for Lifelong Knowledge Editing in LLMs

This paper proposes NMKE, a framework that identifies two categories of knowledge neurons—knowledge-general and knowledge-specific—via neuron-level attribution, and applies entropy-guided dynamic sparse masking to achieve precise neuron-level knowledge editing. NMKE maintains high edit success rates and general model capabilities after 5,000 consecutive edits.

KScope: A Framework for Characterizing the Knowledge Status of Language Models

This paper proposes a five-category taxonomy of LLM knowledge status (Consistent Correct / Conflicting Correct / Missing / Conflicting Incorrect / Consistent Incorrect) and the KScope hierarchical statistical testing framework. By combining repeated sampling with multi-step hypothesis testing, KScope precisely characterizes the modal structure of an LLM's knowledge for a given question, and systematically investigates how context updates each knowledge state. The study finds that constrained context summarization combined with credibility augmentation improves knowledge update success rates by an average of 4.3%.

MemEIC: A Step Toward Continual and Compositional Knowledge Editing

This paper proposes MemEIC, a three-tier framework for continual and compositional knowledge editing in large vision-language models (LVLMs), combining an external dual-modal retrieval memory (Mem-E), an internal modality-decoupled LoRA adapter (Mem-I), and a brain-inspired Knowledge Connector. MemEIC substantially outperforms existing methods on the newly introduced CCKEB benchmark.

MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs

MEMOIR introduces a framework that incorporates zero-initialized residual memory matrices into FFN layers, employs TopHash-based sparse masks to confine each edit to a distinct subset of memory parameters, and at inference time conditionally activates stored knowledge by measuring mask overlap. The approach achieves an optimal balance among reliability, generalization, and locality across 15,000 sequential edits.

Rethinking Residual Distribution in Locate-then-Edit Model Editing

This paper reveals that the residual distribution mechanism in locate-then-edit model editing introduces weight deviation errors that grow with distribution distance, batch size, and sequential edit length. It proposes BLUE (Boundary Layer UpdatE), a strategy that updates only the first and last critical layers, achieving an average improvement of 35.59%.

UniEdit: A Unified Knowledge Editing Benchmark for Large Language Models

This paper presents UniEdit — the first unified LLM knowledge editing benchmark built upon an open-domain knowledge graph (Wikidata), covering 311K samples across 25 domains in 5 major categories. By introducing the Neighborhood Multi-hop Chain Sampling (NMCS) algorithm, UniEdit integrates diverse generalization and locality evaluation criteria into a single framework, systematically revealing the shortcomings of existing editing methods under complex ripple effect evaluations.