微信内可能无法直接打开本站。请点右上角 ··· → 在浏览器打开,或复制链接。
KLOD: Locality-Preserving Knowledge Editing via Non-Target Distribution Preservation
RSS 官方收录 · 可信分层展示
关键摘要
arXiv:2608.…
- 27839v1 Announce Type: new Abstract: Fine-tuning-based knowledge editi…
- In sequential editing, such unconstrained redistribution can accumulat…
- We propose KLOD, a bounded and distribution-preserving objective for f…
摘要引擎:抽取
正文提要
arXiv:2608.27839v1 Announce Type: new Abstract: Fine-tuning-based knowledge editing is simple and architecture-agnostic, but standard cross-entropy increases the edited target probability without explicitly constraining changes in the non-target output distribution. In sequential editing, such unconstrained redistribution can accumulate as distributional drift and contribute to locality degradation. We propose KLOD, a bounded and distribution-preserving objective for fine-tuning-based knowledge editing that separates the intended target update from distributions that should remain stable. KLOD stops target amplification once a probability threshold is reached, while preserving the target-excluded non-target distribution at target positions and the full next-token distribution at prefix positions. Experiments on CounterFact and ZsRE with Llama3-8B-Instruct and Qwen2.5-7B-Instruct show that KLOD substantially mitigates locality degradation while maintaining high edit reliability. The target probability threshold further provides a controllable Generalization--Locality trade-off. Ablation, multi-seed, and distributional KL analyses support the interpretation that KLOD's locality gains are associated with preserving output distributions rather than simply weakening the edit. Code is available on GitHub https://github.com/Hostoday/KLOD .