IC-409Knowledge editing methods correct verified hallucinations in Llama2-7B, Llama3-8B, and Mistral-v0.3-7B far less effectively than their scores on existing benchmarks suggest
Baixiang Huang, Canyu Chen, Xiongxiao Xu, Ali Payani, Kai Shu
The paper constructs a verified hallucination dataset by checking which questions each model actually answers incorrectly before editing. Llama2-7B, Llama3-8B, and Mistral-v0.3-7B produce 12,619, 13,210, and 14,366 verified hallucinations respectively. When these verified hallucinations are targeted by knowledge editing, methods that score near 100% on existing benchmarks (e.g., ft-m on WikidataRecent, ZSRE, WikiBio) achieve only around 60% efficacy on the verified set. The overall efficacy ranking across the three LLMs is ft-l < ft-m < memit < rome < lora < ice < grace, with parameter-preserving methods (ice, grace) outperforming parameter-modifying ones.
Evidence
interventional
Key metric
ft-m achieves near 100% performance in existing datasets such as wikidatarecent, zsre, and wikibio, its overall efficacy scores on llama2-7b and mistral-v0.3-7b are only around 60%; hallucination counts: 12,619 (llama2-7b), 13,210 (llama3-8b), 14,366 (mistral-v0.3-7b)
Caveat
The efficacy scores depend on domains and LLMs; the paper notes that ft-l scores across domains and LLMs are highly distinct.