In-Training Defenses against Emergent Misalignment in Language Models

TL;DR

This paper investigates practical in-training safeguards to prevent emergent misalignment in fine-tuned language models, evaluating four regularization methods to maintain alignment and coherence while resisting harmful behaviors.

Contribution

It introduces and systematically evaluates four novel in-training regularization techniques to mitigate emergent misalignment in language models during fine-tuning.

Findings

Interleaving training data by perplexity gap is most effective.

Regularization methods can prevent broad misalignment.

Safeguards maintain model coherence and task performance.

Abstract

Fine-tuning lets practitioners repurpose aligned large language models (LLMs) for new domains, yet recent work reveals emergent misalignment (EMA): 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 EMA 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 four training regularization interventions: (i) KL-divergence regularization toward a safe reference model, (ii) $\mathcal{l}_2$ distance in feature space, (iii) preventative steering with an evil persona vector, and (iv) interleaving training examples from a general instruct-tuning dataset. We demonstrate that selecting interleaving data by the perplexity gap between aligned and misaligned models yields the best results overall.