Self-Recognition Finetuning can Reverse and Mitigate Emergent Misalignment (Revision)

Language models exhibit metacognitive behaviors such as confidence reporting, self-recognition, and situational awareness. While these behaviors have been studied to understand cognitive capabilities, their interaction effects with downstream alignment remain underexplored. We investigate the relationship between self-recognition i.e a model’s ability to identify its own outputs and emergent misalignment (EM), where narrow fine-tuning produces broadly misaligned behavior. Our central result shows that Self-Generated Text Recognition (SGTR) fine-tuning can both mitigate EM when applied beforehand and reverse it when applied afterward. Our findings hold across three models (GPT-4.1, Qwen2.5-32B-Instruct, Seed-OSS-36B-Instruct) and three EM datasets. We further establish a causal relationship by showing that non-metacognitive baselines don’t produce similar defensive effects, that EM depends critically on identity system prompts embedded in chat templates for Qwen2.5-32B-Instruct, and that
direct Identity Confusion finetuning can further exacerbate the effects of EM finetuning. These
results combined suggest that interventions targeting LLM self-recognition offer a promising
approach to controlling emergent misalignment that does not require knowledge of the EM dataset domain or access to model internals.