Reproduce “Linear Personality Probing and Steering in LLMs: A Big Five Study” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Michel Frising, Daniel Balcells
The paper learns activation directions from Big Five questionnaire scores generated for 406 fictional characters and adds them to Llama 3.3 70B at inference time. It tests whether these directions separate trait-related adjectives and shift choices in a forced-choice extraversion questionnaire. The directions shift selections monotonically within the tested range, but character descriptions remove the effect, limiting claims about general personality control.
Key insights
- Big Five personality traits are encoded in specific, trait-aligned linear subspaces within an LLM's activation space, distinct from general directions of maximal variance.
- These regression-derived linear directions generalize effectively to cleanly separate trait-relevant adjectives, indicating that middle to late layers of the model capture pertinent semantic structure.
- The effectiveness of linear steering is highly context-dependent, with explicit textual personality cues in prompts consistently overriding vector-based interventions in both forced-choice and open-ended tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.