Plan B: Training LLMs to fail less severely

Julian Stastny, Niels Warncke, Dylan Xu, Aengus Lynch, Fazl Barez, Henry Sleight, Ethan Perez

Keywords: AI safety, data poisoning, alignment, robustness, sleeper agents, model organisms, jailbreaks

Abstract: Safety-trained LLMs can produce harmful responses across various input types, as shown by research on jailbreaks, data poisoning, and misalignment. Despite ongoing efforts, fully preventing such failures remains difficult. In this work, we propose a second line of defense: instead of solely focusing on eliminating harmful responses, we also aim to reduce their severity when they occur. As a case study, we experiment with an LLM trained to respond to a backdoor-trigger by complying with harmful requests. We fine-tune the model, without using the trigger in the training data, on the following pairwise preferences: (1) refusal is preferred over any harmful response, (2) less harmful responses are preferred over more harmful ones. We find that training on this preference ordering significantly reduces the harmfulness of backdoor-triggered responses. Finally, we demonstrate that our approach generalizes to several state-of-the-art jailbreak techniques.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3598