Research map / Large language models / Instruction and fine-tuning
Neuron based Personality Trait Induction in Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NPTI, which identifies personality-related FFN neurons in LLMs via activation-probability differences between opposing trait aspects (using the new PersonalityBench dataset) and steers personality by scaling those neurons, without weight updates.
Problem
Inducing and controlling Big Five personality traits in LLMs is currently done via prompting (unstable) or training (costly); existing neuron/head-based methods lack psychological grounding or fine-grained control.
Approach
Identify personality neurons by contrasting activation probabilities between opposite aspects of a trait (e.g., extraverted vs introverted) on a generative benchmark, then amplify positive-aspect neurons and deactivate opposing-aspect neurons at inference, with a sigmoid-weighted scaling by activation difference.
Key results
- On LLaMA-3-8B-Instruct, NPTI achieves average trait mean 9.43 with variance 0.49, matching Simple Prompt and P2 (9.43) and close to LoRA SFT (9.61), with lower variance than prompt baselines.
- Human evaluation gives NPTI the top overall average rank (2.27) versus SFT 2.37, P2 2.32, Simple Prompt 2.79, PAS 3.40.
- General benchmark drops are small for most neuron settings (e.g., GSM8K -1.1 for agreeableness+), with conscientiousness+ slightly improving GSM8K, IFEval and CommonsenseQA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NPTI instead of SFT | LLaMA-3-8B-Instruct · PersonalityBench (average) | personality score mean: 9.61 vs 9.43 | Hurt |
| Use NPTI instead of P² | LLaMA-3-8B-Instruct · PersonalityBench human evaluation (average) | average rank: 2.32 -> 2.27 | Helped |
| Use NPTI instead of Simple Prompt | LLaMA-3-8B-Instruct · PersonalityBench fluency (average) | fluency mean: 9.94 -> 9.86 | Hurt |
| Activate positive conscientiousness neurons | LLaMA-3-8B-Instruct · GSM8K, IFEval (loose), IFEval (strict), CommonsenseQA | benchmark score: +0.3 GSM8K; +0.4 IFEval(loose); +0.4 IFEval(strict); +0.6 CommonsenseQA | Helped |
| Activate positive neuroticism neurons | LLaMA-3-8B-Instruct · CommonsenseQA | score: ↓ 7.0 | Hurt |
| Activate positive neuroticism neurons | LLaMA-3-8B-Instruct · GSM8K, IFEval (loose), IFEval (strict) | benchmark score: ↓ 2.4 GSM8K; ↓ 3.7 IFEval(loose); ↓ 4.7 IFEval(strict) | Hurt |
| Remove the weighted function from neuron manipulation (ablation) | NPTI · PersonalityBench | fluency | Hurt |
| Remove the weighted function from neuron manipulation (ablation) | NPTI · PersonalityBench | personality score | No clear effect |
| Increase γ (magnitude of neuron modification) (ablation) | NPTI · Agreeableness | personality score | Helped |
| Increase γ (magnitude of neuron modification) (ablation) | NPTI · Agreeableness | fluency score | Hurt |
| Lower the neuron activation-difference threshold, selecting more neurons (ablation) | NPTI · Agreeableness | personality score | Helped |
| Lower the neuron activation-difference threshold, selecting more neurons (ablation) | NPTI · Agreeableness | fluency score | Hurt |
Surprising findings
- Activating conscientiousness-positive neurons slightly improved performance on GSM8K, IFEval, and CommonsenseQA, while neuroticism-positive activation caused the largest drop (CommonsenseQA -7.0).
Limitations
- Human evaluation involves only 200 questions and five annotators; the automatic judge is ChatGPT.
- Neuron manipulation reduces general capabilities for some traits (e.g., neuroticism+ with CommonsenseQA -7.0), and the hyperparameter gamma must be tuned per model.
- Evaluation relies on a relatively small test set (~450 instances) derived from SocialIQA.
Experimental setup
Datasets: PersonalityBench (constructed from IPIP-NEO-300 facets and UltraChat topics) SocialIQA GSM8K IFEval CommonsenseQA
Benchmarks: PersonalityBench (SocialIQA-based evaluation split) GSM8K IFEval CommonsenseQA
Metrics: personality trait score (1-5) variance fluency score average human rank accuracy
Baselines: Simple prompt induction P2 PAS ActAdd SFT (LoRA)
Builds on: Big Five personality traits IPIP-NEO-300 P2 personality induction ActAdd LoRA
Closest papers by meaning
- Personality Alignment of Large Language Models (ICLR 2025 · similarity 0.87)
- Open Models, Closed Minds? On Agents Capabilities in Mimicking Human Personalities through Open Large Language Models (AAAI 2025 · similarity 0.86)
- The Language Model Can Have the Personality: Joint Learning for Personality Enhanced Language Model (Student Abstract) (AAAI 2024 · similarity 0.86)
- PICLe: Eliciting Diverse Behaviors from Large Language Models with Persona In-Context Learning (ICML 2024 · similarity 0.86)
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models (NeurIPS 2026 · similarity 0.86)
- LLM vs Small Model? Large Language Model Based Text Augmentation Enhanced Personality Detection Model (AAAI 2024 · similarity 0.85)
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.