Study says the original knowledge of edited models can be recovered via prompting

1. 12:01
   arXiv

   The Truth Beneath the Tokens: Prompt-Space Recovery of Edited Knowledge in Language Models

   The researchers propose Memento, a mechanism that audits the residual knowledge of edited models through prompt-space intervention, guiding an edited model to recover its pre-edit answers without modifying model weights. The method learns a reusable recovery signal, supporting both continuous recovery prefixes and discrete text prompts, and is applicable to standard text interfaces. In experiments on three LLMs and five editing techniques, Memento recovered 80.07% of held-out edited facts, a relative improvement of 9.58% over the baseline, indicating that editing may suppress facts but not fully eliminate their recoverability. On unedited original LLMs, Memento relatively improves average fact recall by 53.9% over standard inference. The authors state that these results demonstrate the potential of prompt-space recovery for auditing residual knowledge in edited models and for studying the gap between factual knowledge and its accessibility at inference time, and that they will release the source code.

Current attention 34·Peak in the comparable period 37 (Nov 10 22:00)·Change over the comparable past 24 hours 0%