![[2602.15847] Do Personality Traits Interfere? Geometric Limitations of Steering in Large Language Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.15847] Do Personality Traits Interfere? Geometric Limitations of Steering in Large Language Models
Summary
This article explores the geometric limitations of steering personality traits in large language models (LLMs), revealing that traits are interdependent, challenging the assumption of independent control.
Why It Matters
Understanding the geometric relationships between personality traits in LLMs is crucial for improving their design and application. This research highlights the complexities of steering traits, which can impact user interactions and model performance in real-world applications.
Key Takeaways
- Personality traits in LLMs exhibit geometric dependence, affecting independent control.
- Steering one personality trait can inadvertently alter others.
- Hard orthonormalisation may enforce independence but reduces steering strength.
- The findings challenge existing assumptions about personality steering in AI.
- Understanding these limitations is vital for future LLM development.
Computer Science > Computation and Language arXiv:2602.15847 (cs) [Submitted on 23 Jan 2026] Title:Do Personality Traits Interfere? Geometric Limitations of Steering in Large Language Models Authors:Pranav Bhandari, Usman Naseem, Mehwish Nasim View a PDF of the paper titled Do Personality Traits Interfere? Geometric Limitations of Steering in Large Language Models, by Pranav Bhandari and 2 other authors View PDF HTML (experimental) Abstract:Personality steering in large language models (LLMs) commonly relies on injecting trait-specific steering vectors, implicitly assuming that personality traits can be controlled independently. In this work, we examine whether this assumption holds by analysing the geometric relationships between Big Five personality steering directions. We study steering vectors extracted from two model families (LLaMA-3-8B and Mistral-8B) and apply a range of geometric conditioning schemes, from unconstrained directions to soft and hard orthonormalisation. Our results show that personality steering directions exhibit substantial geometric dependence: steering one trait consistently induces changes in others, even when linear overlap is explicitly removed. While hard orthonormalisation enforces geometric independence, it does not eliminate cross-trait behavioural effects and can reduce steering strength. These findings suggest that personality traits in LLMs occupy a slightly coupled subspace, limiting fully independent trait control. Subjects: Computati...
