Skip to content

โœ๏ธ Knowledge Editing

๐Ÿง  NeurIPS2026 ยท 8 paper notes

๐Ÿ“Œ Same area in other venues: ๐Ÿ“ท CVPR2026 (2) ยท ๐Ÿ”ฌ ICLR2026 (15) ยท ๐Ÿ’ฌ ACL2026 (10) ยท ๐Ÿงช ICML2026 (8) ยท ๐Ÿค– AAAI2026 (4) ยท ๐Ÿง  NeurIPS2025 (6)

๐Ÿ”ฅ Top topics: LLM ร—2

DynEdit: Dynamic Entropy-Guided Sequential Editing for Large Language Models

DynEdit overcomes the rigidity and gradient dilution of fixed-chunk long-form knowledge editing by leveraging pre-edit token-level Shannon entropy to dynamically shrink sliding strides across dense factual regions and adaptively weight token optimization.

Exposing the Illusion of Erasure in Knowledge Editing for LLMs

Framing knowledge editing through adversarial reverse-engineering, this work proves that localized low-rank updates merely push pre-trained facts into complementary null spaces with surface-level suppression logits; universal adversarial suffixes reliably recover supposed "erased" knowledge (>85% guided, up to 48.5% blind), exposing an illusion of erasure.

Golden Layers and Where to Find Them: Improved Knowledge Editing for Large Language Models Via Layer Gradient Analysis

This paper demonstrates the existence of sample-invariant "Golden Layers" that achieve near-optimal knowledge editing across LLMs, and introduces Layer Gradient Analysis (LGA)โ€”a fast, first-order gradient attribution method that accurately pinpoints golden layers without requiring actual editing runs.

MetaKE: Meta-Learning for Knowledge Editing Toward a Better Accuracy-Editability Trade-off

MetaKE bridges the long-standing disconnect between upstream target representation optimization and downstream constrained parameter optimization by formulating knowledge editing as a bi-level optimization problem, using a closed-form Structural Gradient Proxy to provide downstream-aware feedback for target refinement.

Norm Anchors Make Model Edits Last

Reveals that sequential Locate-and-Edit (L&E) failure stems from a self-reinforcing positive norm-feedback loop between edited MLP weights and solved value vectors, and proposes Norm-Anchor Scaling (NAS) to rescale target values to base-model reference norms, extending the editing horizon by over 4ร—.

Open-Vocabulary Domain Unlearning

The paper extends visual-domain unlearning from failure on training classes to failure on classes absent from unlearning training, combining Fisher-masked visual-parameter updates with Targeted Manifold Scattering (TMS) to improve the forgettingโ€“retention trade-off in few-shot CLIP classification, without establishing certified data deletion or guaranteed semantic preservation.

Steer2Edit: From Activation Steering to Component-Level Editing

Steer2Edit introduces a theoretically grounded, training-free framework with closed-form solutions that converts activation steering vectors into component-level rank-1 parameter edits on attention heads and MLP neurons, yielding superior attribute-utility trade-offs across safety alignment, truthfulness promotion, and reasoning efficiency control.

Synthetic Worlds for Temporal Evaluation and Knowledge Updating in LLMs

This paper introduces ParallelEvents, a contamination-resistant future-world benchmark ensuring local and global consistency, and Synapse, a synthetic training framework combining long-form continued pretraining with preference-steered editing to internalize evolving multi-hop and causal knowledge.