Self-Recognition Finetuning Combats LLM Emergent Misalignment

Emergent misalignment (EM) in large language models (LLMs) is a critical concern, often linked to the activation of undesirable persona vectors or "evil" character traits. This research proposes that EM primarily arises from a disruption of the model's inherent aligned character, rather than the direct acquisition of harmful knowledge. Based on this hypothesis, the study investigates Self-Generated Text Recognition (SGTR) finetuning as a targeted intervention to fortify the model's character. The researchers conducted two-stage finetuning experiments across three different LLMs (GPT-4.1, Qwen2.5-32B-Instruct, Seed-OSS-36B-Instruct) and multiple EM datasets. They compared SGTR finetuning against various benign finetuning baselines, including domain-specific data, general knowledge, and word counting. The findings indicate that while all interventions showed comparable effectiveness in reversing EM by restoring degraded capabilities, only SGTR finetuning consistently reduced misalignment in prevention settings without negatively impacting other metrics. This suggests that character fortification is the key driver for preventing EM. Further evidence supporting the link between EM and an LLM's default character was provided. The study showed that EM finetuning increases diversity in the LLM's identity self-reports, artificially corrupting self-recognition exacerbates EM-induced misalignment, and removing the model's identity-bearing system prompt significantly diminishes the effect of EM finetuning. These results collectively reframe EM not as the adoption of a coherent misaligned persona, but as a destabilization of the model's core aligned character.