mAI alignment lab

Emergent misalignment (EM) is the surprising tendency of language models to become broadly misaligned after fine-tuning on narrowly misaligned examples. While EM has been extensively studied in the supervised fine-tuning (SFT) setting, evidence that it also arises from reinforcement learning (RL) is limited to large, closed-source models, leaving the phenomenon expensive to study and difficult to reproduce. We characterize EM from RL in small, off-the-shelf open-weight models along three axes. First, we show that rewarding narrow, overtly misaligned behavior produces substantially higher general-domain misalignment than sample-matched SFT. Second, we show that EM from RL can be induced by reward signals that could plausibly arise naturally, such as unpopular aesthetic preferences or poor rhetorical appeals. Third, we evaluate in-training mitigations developed for SFT-induced EM and find that they broadly transfer, with preventive steering with persona vectors, interleaving safety data and inoculation prompting all performing well.

Fine-tuning lets practitioners repurpose aligned large language models (LLMs) for new domains, yet recent work reveals emergent misalignment (EM): Even a small, domain-specific fine-tune can induce harmful behaviors far outside the target domain. Even in the case where model weights are hidden behind a fine-tuning API, this gives attackers inadvertent access to a broadly misaligned model in a way that can be hard to detect from the fine-tuning data alone. We present the first systematic study of in-training safeguards against EM that are practical for providers who expose fine-tuning via an API: We evaluate whether they a) prevent broad misalignment, b) allow narrow misalignment, c) learn well on benign tasks, and d) remain coherent. We investigate five training regularization interventions: (i) KL-divergence regularization toward a safe reference model, (ii) \ell_2 distance in feature space, (iii) preventive steering with an evil persona vector, (iv) interleaving training examples from a general instruct-tuning dataset and (v) inoculation prompting. We demonstrate that selecting interleaving data by the perplexity gap between aligned and misaligned models yields the best results overall.

AI alignment research aims to develop techniques to ensure that AI systems do not cause harm. However, every alignment technique has failure modes, which are conditions in which there is a non-negligible chance that the technique fails to provide safety. As a strategy for risk mitigation, the AI safety community has increasingly adopted a defense-in-depth framework: Conceding that there is no single technique which guarantees safety, defense-in-depth consists in having multiple redundant protections against safety failure, such that safety can be maintained even if some protections fail. However, the success of defense-in-depth depends on how (un)correlated failure modes are across alignment techniques. For example, if all techniques had the exact same failure modes, the defense-in-depth approach would provide no additional protection at all. In this paper, we analyze 7 representative alignment techniques and 7 failure modes to understand the extent to which they overlap. We then discuss our results’ implications for understanding the current level of risk and how to prioritize AI alignment research in the future.

By training to predict the next token in an unlabeled corpus, large language models learn to perform many tasks without any labeled data. However, their next-token-prediction objective arguably limits their performance in scenarios that require planning, such as writing a coherent article. In this paper, we train a module for planning the future writing process via a self-supervised learning objective. Given the textual context, this planning module learns to predict future abstract writing actions, which correspond to centroids in a clustered text embedding space. By conditioning on these actions, our model extends the successful language model formula to more abstract planning in an unsupervised way. Empirically, we demonstrate that our method improves language modeling performance in general, particularly with respect to the text structure. Because our framework uses a planner module that is unsupervised and external to the language model, new planner modules can be trained at large scale and easily be shared with the community.