✏️ Knowledge Editing

🧪 ICML2026 · 2 paper notes

📌 Same area in other venues: 💬 ACL2026 (4) · 📷 CVPR2026 (2) · 🔬 ICLR2026 (8) · 🤖 AAAI2026 (5) · 🧠 NeurIPS2025 (6)

CrispEdit: Low-Curvature Projections for Scalable Non-Destructive LLM Editing

Formulates LLM editing as "minimize edit loss s.t. capability loss unchanged" constrained optimization, converts it via Bregman divergence equivalence to low-curvature subspace projection using the Gauss-Newton Hessian, and leverages K-FAC plus a Kronecker basis trick that avoids explicit projector construction. This enables 3000 edits on A40 in 6 minutes, while keeping LLaMA-3-8B's average drop on MMLU/IFEval/ARC-C/TruthfulQA/GSM8K under 1%, significantly outperforming AlphaEdit / MEMIT / fine-tuning.

KORE: Enhancing Knowledge Injection for Large Multimodal Models via Knowledge-Oriented Controls

KORE injects new knowledge into LMMs via a two-stage "knowledge-oriented control": it automatically expands a single fact into structured multi-turn dialogues and instruction tasks (improving generalization), while initializing LoRA adapters in the null space of the covariance matrix of prior knowledge (minimizing interference with old capabilities). This achieves both strong adaptation and retention on LLaVA-v1.5 / Qwen2.5-VL.