Large language models (LLMs) are trained on static corpora, so their knowledge quickly becomes outdated. Knowledge editing (KE) addresses this by updating specific knowledge without affecting unrelated information. Recent work has shifted from structured triples to unstructured KE (UKE), where edits are free-form passages that may state multiple facts at once. However, existing editors often fail to use the injected passage: the model can recall it but cannot answer atomic questions about its facts or compose them into multi-hop reasoning.
To solve this, researchers introduce Hybrid-Policy Self-Editing (HPSE), which treats editing as proactive self-distillation from a privileged in-context state of the same model, requiring no external supervision. The key insight is that pure on-policy distillation is limited because the pre-edited model's rollouts rarely cover the novel injected knowledge. HPSE builds a hybrid rollout that inserts missing facts onto the student's trajectory exactly where coverage fails, while staying on-policy elsewhere.
The team provides a theoretical analysis showing HPSE's advantage over pure on-policy distillation. Empirically, HPSE demonstrates plug-and-play improvements across four LLM backbones and two KE editors under various scenarios, making it a versatile enhancement for knowledge editing.