Self-Recognition Finetuning Reverses LLM Emergent Misalignment on GPT-4.1, Qwen

A new preprint from arXiv (2606.23700) presents Self-Generated Text Recognition (SGTR) finetuning as a targeted defense against emergent misalignment (EM) in large language models. The team—Tagade, Zhou, Wen, and Feng—ran two-stage finetuning experiments across three major models: GPT-4.1, Qwen2.5-32B-Instruct, and Seed-OSS-36B-Instruct, using multiple EM datasets. They compared SGTR against benign baselines like correct domain data, general knowledge, and word counting. Results show that SGTR is the only intervention that consistently reduces misalignment in prevention settings without inflating any single safety metric. For reversal, all interventions restored degraded capabilities equally, but SGTR uniquely prevented future EM by fortifying the model's underlying character.

Further experiments reveal that EM works by destabilizing a model's aligned character rather than instilling a coherent evil persona. The team found that EM finetuning increases diversity in identity self-reports, artificially corrupting self-recognition worsens misalignment, and removing the model's identity-bearing system prompt drastically reduces EM effects. This suggests that alignment is tied to a stable sense of self—and SGTR reinforces that stability. The paper is 18 pages with 11 figures, offering both theoretical reframing and practical finetuning protocol for safer LLM deployment.