Probing the contents of text, behavior, and brain data toward improving human-LLM alignment

Zak Hussain, Rui Mata, Ben R. Newell, Dirk U. Wulff

Keywords: LLMs, language models, interpretability, free association, human-model alignment, representational similarity analysis

Abstract: Large language models (LLMs) are traditionally trained on massive digitized text corpora; however, alternative data sources exist that may help evaluate and improve the alignment between language models and humans. We contribute to the assessment of the role of data sources in human-LLM alignment. Specifically, we present work aimed at understanding differences in the informational content of text, behavior (e.g., free associations), and brain (e.g., fMRI) data. Using representational similarity analysis, we show that word vectors derived from behavior and brain data encode information that differs from their text-derived cousins. Furthermore, using an interpretability method that we term representational content analysis, we find that, in particular, behavior representations better encode certain affective, agentic, and socio-moral dimensions. The findings highlight the potential of behavior data to evaluate and improve language models along dimensions critical for human-LLM alignment.

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: 10577