✏️ Knowledge Editing

🤖 AAAI2026 · 5 paper notes

Catastrophic Forgetting in Kolmogorov-Arnold Networks

The first systematic study of catastrophic forgetting in Kolmogorov-Arnold Networks (KANs): establishes a theoretical framework linking forgetting to activation support overlap and intrinsic data dimensionality, and proposes KAN-LoRA for continual fine-tuning knowledge editing in language models.

Hybrid-DMKG: A Hybrid Reasoning Framework over Dynamic Multimodal Knowledge Graphs for Multimodal Multihop QA with Knowledge Editing

This paper proposes the MMQAKE benchmark and the Hybrid-DMKG framework, which constructs a dual-channel hybrid reasoning mechanism — combining relation link prediction with RAG-augmented LVLM inference — over a dynamic multimodal knowledge graph, supplemented by a background reflection decision module. The approach significantly outperforms existing methods on 2–5 hop multimodal knowledge editing QA (H-Acc of 29.90% on LLaVA, surpassing IKE by 13.52 percentage points).

Is the Information Bottleneck Robust Enough? Towards Label-Noise Resistant Information Bottleneck Learning

This paper identifies the inherent vulnerability of the Information Bottleneck (IB) principle under label noise and proposes LaT-IB, which decomposes representations into a clean-label subspace and a noisy-label subspace. Combined with a Minimal-Sufficient-Clean (MSC) criterion and a three-stage training framework, LaT-IB significantly outperforms existing IB methods across diverse noise conditions.

Model Editing as a Double-Edged Sword: Steering Agent Ethical Behavior

This paper frames the steering of agent ethical behavior as a model editing task (Behavior Editing), proposes a three-tier BehaviorBench grounded in psychological moral theory, and validates on 9 open-source and 20 closed-source models that model editing can precisely steer agents toward either benevolent or malicious behavior, with a single edit potentially causing global moral alignment drift.

Multiplicative Orthogonal Sequential Editing for Language Models (MOSE)

This paper proposes MOSE (Multiplicative Orthogonal Sequential Editing), which injects new knowledge by left-multiplying the parameter matrix with an orthogonal matrix (rather than via additive updates), strictly preserving the Frobenius norm and condition number of the edited matrix. MOSE achieves a 12.08% performance improvement in sequential editing while retaining 95.73% of general capabilities.