knowledge-editing

FedAlphaEdit 提出了首个在本地编辑与服务器端合并规则上统一遵循零空间原则的协作式知识编辑框架，使无法共享原始编辑数据的机构能够联合维护共享模型，其编辑成功率与知识保留能力接近集中式编辑水平。

EngramEdit 提出一种通过条件记忆（conditional memory）实现 LLM 解耦知识更新的方法，联合更新共享 n-gram 嵌入以实现近乎完美的编辑成功率，在多跳推理中达到最强基线近三倍的准确率，同时保留无关知识与通用能力。

The paper introduces FOVEATED, a plug-and-play framework that improves atomic-fact recall in unstructured knowledge editing by perturbing RoPE positions of preceding context during editing to counteract context-induced difficulty underestimation, showing consistent gains across five editors, two LLM backbones, and three benchmarks.

MedKIT is a new benchmark evaluating how well LLMs integrate and apply updated clinical knowledge across lexical, relational, compositional, and operational probes. A large-scale study of 12 knowledge integration strategies across 5 models reveals a persistent gap between factual recall and usable knowledge, with little improvement on compositional or operational reasoning.

ALOE introduces a semantically addressed low-rank operator for knowledge editing in language models, improving edit scope and precision with high efficacy and locality on standard benchmarks.

This paper proposes a spectral-based framework for selectively reversing knowledge edits in large language models, allowing removal of harmful edits while preserving beneficial ones.

This paper presents a negative result showing that current knowledge-editing benchmarks cannot effectively evaluate scope classifiers, using INLAY, a gradient-free editor, to demonstrate that no per-query routing method can improve performance due to structural limitations in the benchmarks.

DOW-KE introduces an anchor-free method for knowledge editing in large language models by directly optimizing weight updates across layers, achieving higher performance in sequential editing tasks.

This paper introduces HPSE, a hybrid-policy self-editing method for composable unstructured knowledge editing in LLMs, enabling edited models to answer atomic questions and perform multi-hop reasoning about injected knowledge. The approach uses proactive self-distillation with hybrid rollouts and is validated across multiple LLM backbones and editors.

ForgetBench introduces a benchmark to systematically characterize forgetting behavior in large language models under continual knowledge editing, using concept-based and scenario-based QA to measure temporal decay and retention dynamics.

Moir is a method that improves cross-domain knowledge editing in LLMs by aligning the preservation distribution with the model's own decoding distribution, avoiding reliance on external corpora. It consistently preserves complex capabilities like mathematical reasoning across multiple models and editors.

KnowledgeDebugger is a GUI-based tool for exploring knowledge localization and editing in Transformer models, integrating state-of-the-art knowledge editing methods from the EasyEdit library to support intuitive analysis and hypothesis formation.

This paper identifies and addresses the 'editing decoupling failure' in Multimodal LLMs, where knowledge updates via multimodal inputs fail to generalize to unimodal queries. The authors propose DECODE, a method to disentangle and localize modality-specific neurons for more effective knowledge editing.

Introduces a benchmark to evaluate how knowledge editing methods handle logical consequences of fact edits, revealing that existing approaches like ROME and FT accurately insert direct assertions but fail to propagate entailed knowledge, with a performance gap of up to 24%.

HyperPatch proposes a parameter-preserving framework for sequential knowledge editing under n-ary structural drift, using hypergraph neural networks to maintain event integrity. It achieves 96.24% and 21.06% relative improvements in Hop-wise Accuracy on MQuAKE-CF and MQuAKE-T benchmarks, respectively.

This paper introduces the FactualOpinionEditing with Evidence (FOE) benchmark to assess the ability to edit factual opinions in LLMs, and proposes a Self-Generated Evidence-Aligned method to improve opinion-evidence alignment.

This paper presents a theoretical analysis of parameter-based knowledge editing in LLMs, revealing how localized edits can cause global reasoning collapse, and provides empirical evidence that retrieval-based methods outperform parameter-editing approaches.

This paper introduces UniKE, the first benchmark for cross-modal knowledge editing in unified multimodal models (UMMs), revealing a significant modality gap where text edits achieve 92% efficacy but only 18.5% transfer to image generation. It proposes Reasoning-augmented Parameter Editing to improve cross-modal transfer, with gains up to 18.6 percentage points.

This paper investigates the internal mechanisms of knowledge editing methods ROME and MEMIT, revealing that edits rely on a common functional subspace of weights and suppress rather than overwrite knowledge, explaining why edits fail to propagate to related facts.

This paper investigates the mechanisms underlying sequential knowledge editing in LLMs, showing that many regularization strategies are unnecessary and that stability emerges naturally from properly accounting for accumulated editing constraints.