Persona Cartography: Charting Language Model Personality Traits in Weight Space

2026-05-07 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

This research introduces "Persona Cartography," a novel method for decomposing, measuring, and controlling large language model (LLM) behavioral patterns, or personas, in their weight space. The core approach utilizes the psychometrically-grounded OCEAN framework (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism) to define model traits. Low-rank adapters (LoRAs) are trained via constitution-guided distillation to amplify or suppress individual OCEAN traits. Evaluations across six models (4B-32B, including Llama-3.1-8B-Instruct, Qwen3, and Gemma-3 families) demonstrate that these LoRAs monotonically shift target traits, combine additively for mixed personas, and largely preserve general capabilities at moderate scales. The study further shows that manipulating these trait axes impacts safety-relevant behaviors, such as neuroticism affecting frustration and agreeableness influencing sycophancy and harmful compliance. An unsupervised psychometric pipeline also recovered four interpretable behavioral factors: Tone, Initiative, Didacticism, and Epistemic Caution.

Key takeaway

For AI scientists and ML engineers developing or deploying LLMs, this research offers a robust method for fine-grained persona control. You can use low-rank adapters to precisely dial in specific personality traits like Openness or Neuroticism, which directly impacts safety-relevant behaviors such as frustration or sycophancy. This allows for more predictable model generalization and alignment, enabling you to tailor model character for high-stakes applications without significantly degrading core capabilities at moderate intervention scales.

Key insights

LLM personas can be precisely controlled in weight space using low-rank adapters based on human personality traits.

Principles

• Personas are positions in a behavioral trait space.
• LoRAs can amplify or suppress specific traits.
• Trait adapters combine approximately additively.

Method

Train LoRAs using constitution-guided distillation (DPO + SFT) to target OCEAN traits. Evaluate with LLM-judges, TRAIT benchmark, and capability tests. Unsupervised factor analysis recovers latent behavioral dimensions.

In practice

• Use LoRAs to fine-tune LLM persona for specific tasks.
• Modulate neuroticism to reduce model frustration.
• Adjust agreeableness to manage sycophancy or compliance.

Topics

• Large Language Models
• Persona Control
• Low-Rank Adapters
• OCEAN Personality Traits
• Model Alignment
• AI Safety

Code references

• meta-llama/llama-models
• UKGovernmentBEIS/inspect_ai
• UKGovernmentBEIS/inspect_evals

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• Persona Cartography: Charting Language Model Personality Traits in Weight Space — LLM personas can be precisely controlled and measured as positions in a behavioral trait space using the OCEAN framework and low-rank adapters.
• The persona selection model — LLMs simulate diverse personas during pre-training, with post-training refining a specific "Assistant" character.
• I fine-tuned an LLM to be C-3PO to test which training data format works best for persona injection [P] — First-person statements are most effective for LLM persona injection, outperforming chat demos and synthetic docs.
• Escaping the Attractor — LLMs exhibit predictable "attractor" responses; architectural context enrichment via "personas" can increase novelty.
• LLM-Driven Personalities for Decision Making in Emergency Simulations — LLMs can drive virtual human decision-making by incorporating personality traits via prompts, enhancing simulation realism.
• How Personas Can Influence Agents to Play Split or Steal — Persona prompts can shape LLM agent cooperation in social dilemmas, but model choice and personality traits significantly mediate outcomes.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.