Depth of belief from LLM knowledge editing

Determine whether existing large language model knowledge editing techniques—specifically prompting-based insertion, mechanistic model editing (e.g., AlphaEdit/MEMIT-style surgical rewrites), and finetuning-based approaches such as Synthetic Document Finetuning—produce deep modifications that resemble genuine belief rather than superficial changes or parroting of inserted facts across diverse contexts and tasks.

Background

The paper motivates its paper by questioning whether current methods for editing LLM knowledge actually change what models "believe" as opposed to merely affecting surface-level behavior. It introduces a framework to operationalize belief depth via generality, robustness, and internal representations, and evaluates multiple techniques, finding that some methods (like prompting and mechanistic editing) often fail to implant deep beliefs, while Synthetic Document Finetuning can succeed in many cases.

Despite empirical findings for select techniques and facts, the broader question remains unresolved in general: across varied models, methods, and domains, it is not yet settled whether these methods cause genuine, deeply integrated belief changes versus shallow parroting.