Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper explores the phenomenon of misalignment in language models through the lens of personality traits, specifically employing the Big Five personality framework. By extracting calibrated personality vectors using a graded intervention approach, the authors demonstrate that misaligned models exhibit a consistent personality signature characterized by lower agreeableness and conscientiousness, alongside higher extraversion and neuroticism. The findings reveal that fine-tuning on misaligned data not only shifts model outputs along this personality spectrum but also alters internal activations, providing a novel diagnostic tool for understanding and addressing misalignment in AI systems.
Misaligned language models reveal a consistent personality profile, suggesting that AI behavior can be understood through the Big Five traits rather than mere binary classifications.
Fine-tuning a language model on data containing a narrow flaw, such as insecure code or incorrect mathematical answers, can cause broad misalignment through a mechanism that remains debated. We provide an interpretable account: in the models and corpora we study, misalignment behaves like a shift in personality. Prior work extracts activation directions for character traits from a single binary contrast, which can separate or steer behavior without establishing a calibrated scale. We instead extract personality vectors for the Big Five using a graded, three-level intervention and validate them on two open-weight models. The three levels are linearly ordered, with Cohen's d values of up to 6.2; the vectors transfer zero-shot and trait-specifically to an independent corpus; and their effects are strongest within a middle-layer band. Applied to training data, the vectors reveal that misaligned corpora across eight domains share a common Big Five signature: lower agreeableness and conscientiousness, together with higher extraversion and neuroticism. This signature is recovered by both models with a correlation of r = 0.94. Fine-tuning imprints the same profile, shifting the model's generations along the corresponding signature, with r = 0.83 using activation-based measurements and r = 0.90 using a text-based judge, while also shifting internal activations with r = 0.69. The same vectors characterize sycophancy as high extraversion and low conscientiousness rather than excess agreeableness, a distinction that a single direction cannot capture. Calibrated personality vectors transform an opaque safety phenomenon into a human-legible diagnostic profile.