Research map / Large language models / Language, safety and interpretability
Tracing the Persona Circuit: How Large Language Models Encode and Express Character Traits
Guanzheng Qin, Chenghao Sun, Zhining Xie, Xinmei Tian
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| inject mean persona activation into top-30 critical heads (PSE) | Llama-3.2-3B-Instruct · Character-LLM multi-turn Personality | LLM-judge score: 4.47 -> 4.81 | Improved |
| inject mean persona activation into top-30 critical heads (PSE) | Llama-3.2-3B-Instruct · RoleBench EN CUS | score: 18.66 -> 20.90 | Improved |
| apply PSE to random 30 heads instead of critical heads (ablation) | PSE · RoleBench EN CUS | score: 20.90 -> 18.44 | Hurt |
| apply PSE to random 30 heads instead of critical heads (ablation) | PSE · Character-LLM single-turn Hallucination | LLM-judge score: 6.05 -> 5.42 (-10.4%) | Hurt |
| increase number of enhanced heads from 10 to 30 (ablation) | PSE · Character-LLM single-turn Personality | LLM-judge score: 5.49 -> 5.60 | Improved |
| increase number of enhanced heads from 30 to 40 (ablation) | PSE · RoleBench EN CUS | score: 20.90 -> 20.87 | No clear effect |
| add PSE | Llama-3.2-3B-Instruct · MMLU | accuracy: 62.21 -> 61.83 | No clear effect |
| add PSE (top-30) | Llama-3.2-3B-Instruct · GSM8K | accuracy: 59.74 -> 60.80 | No clear effect |
| add representation-engineering steering | Llama-3.2-3B-Instruct · MMLU | accuracy: 62.21 -> 61.38 | Hurt |
| add prompt reminder (Kong et al., 2024) | Llama-3.2-3B-Instruct · Character-LLM Personality | LLM-judge score: 5.14 -> 5.13 (single) | No clear effect |
| add PSE | Llama-3.2-3B-Instruct · Character-LLM Memory (multi-turn) | LLM-judge score: 2.49 -> 2.50 | No clear effect |
| ablate under 5% of identified attention heads (ablation) | Llama-3.2-3B-Instruct · Trait induction (evil) | personality score | Hurt |
Closest papers by meaning
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models (NeurIPS 2026 · similarity 0.91)
- Persona-Model Collapse in Emergent Misalignment (NeurIPS 2026 · similarity 0.88)
- PersonaManifold: Revealing and Exploiting Curved Geometry in LLM Persona Representations (NeurIPS 2026 · similarity 0.87)
- PICLe: Eliciting Diverse Behaviors from Large Language Models with Persona In-Context Learning (ICML 2024 · similarity 0.86)
- Open Models, Closed Minds? On Agents Capabilities in Mimicking Human Personalities through Open Large Language Models (AAAI 2025 · similarity 0.86)
- Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
