Megadose AI progress, ranked and analyzed.

Hybrid-Policy Self-Editing for Composable Unstructured Knowledge Editing

· HF Daily Papers ·
The paper targets a failure mode where edited models can repeat an added passage but cannot use its facts compositionally.

The authors frame unstructured knowledge editing as self-distillation from the same model in a privileged in-context state. Their HPSE method mixes on-policy rollout with targeted insertion of missing facts when the model’s own trajectory fails to cover new knowledge. They report plug-and-play gains across four LLM backbones and two knowledge-editing methods. HF Daily Papers' note

score 5

Categories: Research