LLMs Can Revoke Learned States with Process Sidecars

Large language models often undergo multi-stage adaptation, including initial skill acquisition, private memory integration, and a final safety phase to prevent undesirable outputs. A significant challenge arises when attempting to revoke a specific memory after the safety phase, as the subsequent training can subtly "transport" or alter the memory's influence. Simply subtracting the initial memory update is insufficient. Researchers have developed "process sidecars," a sophisticated editing technique designed to address this problem. This method involves a two-coefficient edit family that leverages information from the future safety-training process itself. By analyzing how the safety optimizer has bent the memory direction, the sidecar approach can more precisely remove the targeted memory. The paper demonstrates that this technique can recover a counterfactual safety-only state with second-order accuracy, outperforming simpler "task arithmetic" edits. Experimental results across various models show improved refusal closure for held-out data, indicating a more effective and robust way to manage and revoke learned states in LLMs.