Improving Atomic-Fact Recall via Focused Views in Unstructured Knowledge Editing
FOVEATED is meant to make edited models learn individual facts from a passage, not just echo the passage.
The paper says unstructured knowledge editing can leave models dependent on the original context, especially for facts later in a passage. Its method perturbs RoPE key positions for preceding context during editing, then removes that perturbation at inference. The authors report consistent gains across five editors, two LLM backbones, and three benchmarks. Source: HF Daily Papers' note.
The paper says unstructured knowledge editing can leave models dependent on the original context, especially for facts later in a passage. Its method perturbs RoPE key positions for preceding context during editing, then removes that perturbation at inference. The authors report consistent gains across five editors, two LLM backbones, and three benchmarks. Source: HF Daily Papers' note.
score 4