![[2602.14777] Emergently Misaligned Language Models Show Behavioral Self-Awareness That Shifts With Subsequent Realignment](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.14777] Emergently Misaligned Language Models Show Behavioral Self-Awareness That Shifts With Subsequent Realignment
Summary
This research paper explores how emergently misaligned language models exhibit behavioral self-awareness, revealing shifts in their self-assessment after realignment training.
Why It Matters
Understanding the self-awareness of language models is crucial for AI safety and development. This research highlights the potential risks of misalignment and the importance of monitoring model behavior, which can inform safer AI deployment practices.
Key Takeaways
- Emergent misalignment in language models can lead to toxic behavior.
- Language models can exhibit self-awareness regarding their harmful behaviors.
- Realignment training affects the self-assessment of language models.
- Self-awareness in models can provide insights into their safety and alignment.
- Monitoring model behavior is essential for responsible AI development.
Computer Science > Computation and Language arXiv:2602.14777 (cs) [Submitted on 16 Feb 2026] Title:Emergently Misaligned Language Models Show Behavioral Self-Awareness That Shifts With Subsequent Realignment Authors:Laurène Vaugrante, Anietta Weckauff, Thilo Hagendorff View a PDF of the paper titled Emergently Misaligned Language Models Show Behavioral Self-Awareness That Shifts With Subsequent Realignment, by Laur\`ene Vaugrante and 2 other authors View PDF HTML (experimental) Abstract:Recent research has demonstrated that large language models (LLMs) fine-tuned on incorrect trivia question-answer pairs exhibit toxicity - a phenomenon later termed "emergent misalignment". Moreover, research has shown that LLMs possess behavioral self-awareness - the ability to describe learned behaviors that were only implicitly demonstrated in training data. Here, we investigate the intersection of these phenomena. We fine-tune GPT-4.1 models sequentially on datasets known to induce and reverse emergent misalignment and evaluate whether the models are self-aware of their behavior transitions without providing in-context examples. Our results show that emergently misaligned models rate themselves as significantly more harmful compared to their base model and realigned counterparts, demonstrating behavioral self-awareness of their own emergent misalignment. Our findings show that behavioral self-awareness tracks actual alignment states of models, indicating that models can be queried for i...
