LLM persona unlearning

The researchers propose the large language model persona unlearning task, which uses weight-level editing to make a specified persona difficult to elicit and express in unseen contexts; the motivation is that runtime control of open-weight models may be removed. The team releases PersonaUnlearnBench, covering six models from three families and five personas, including aligned forget/retain sets, held-out instruction rewrites, and four-axis evaluation; the benchmark shows that standard unlearning methods struggle to reliably erase the target persona without sacrificing meaningful generation or general capabilities. The authors propose PaCE, which locates an internal behavior direction by comparing the target response and the ideal response on the same question, then trains the state under the target prompt to deviate from the target pattern and move closer to the ideal response. Experiments show that PaCE can persistently suppress the target persona, maintain relatively high response quality and useful corresponding behavior, with a moderate utility cost.