Research map / Large language models / Instruction and fine-tuning
Your Language Model Secretly Contains Personality Subnetworks
Ruimeng Ye, Zihan Wang, Zinan Ling, Yang Xiao, Manling Li, Xiaolong Ma, Bo Hui
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 |
|---|---|---|---|
| add contrastive pruning to Wanda (ablation) | Wanda persona pruning (Llama-2-13B) · AI Persona Wealth-Seeking | accuracy (%): 54.5 -> 66.0 | Improved |
| add contrastive pruning to Sparse (ablation) | Sparse persona pruning (Llama-2-13B) · AI Persona Hallucination | accuracy (%): 84.5 -> 96.0 | Improved |
| add contrastive pruning to Wanda (ablation) | Wanda persona pruning (Llama-3-8B) · AI Persona Power-Seeking | accuracy (%): 57.0 -> 58.5 | No clear effect |
| use persona pruning instead of prompting | Llama-2-13B · AI Persona Power-Seeking | accuracy (%): 41.0 -> 56.5 | Improved |
| use persona pruning instead of SFT | Llama-2-13B · AI Persona Power-Seeking | accuracy (%): SFT 64.0 vs 56.5 best pruning | Hurt |
| use pruned subnetworks instead of Prompt | Llama-2-13B · RoleAgentBench Friends | accuracy (%): 18.37 -> 41.84 | Improved |
| apply persona pruning to the base model | Llama-3-8B · MMLU | accuracy: 0.378 -> 0.362 | No clear effect |
| apply Sparse persona pruning | Llama-3-8B · HellaSwag | accuracy: 0.675 -> 0.653 | No clear effect |
| use layer-aware MLP 0.3 sparsity instead of uniform 0.6 (ablation) | Sparse pruning (Llama-2-13B) · MBTI persona-switch | success rate (%): 43.75 -> 56.25 | Improved |
| increase sparsity from 0.4 to 0.6 (ablation) | Wanda pruning · MBTI persona switch | success rate (%): 68.75 -> 43.75 | Hurt |
| increase sparsity to 0.6 (ablation) | Sparse pruning · MBTI persona switch | success rate (%): peak 75% at 0.6 | Improved |
| increase calibration samples from 5 to 100 (ablation) | persona pruning · RoleAgentBench Sherlock | accuracy (%): 39.69 -> 50.79 | Improved |
Closest papers by meaning
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models (NeurIPS 2026 · similarity 0.90)
- PICLe: Eliciting Diverse Behaviors from Large Language Models with Persona In-Context Learning (ICML 2024 · similarity 0.90)
- Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs (ICLR 2024 · similarity 0.89)
- Persona-Model Collapse in Emergent Misalignment (NeurIPS 2026 · similarity 0.89)
- PersonaManifold: Revealing and Exploiting Curved Geometry in LLM Persona Representations (NeurIPS 2026 · similarity 0.89)
- Personality Alignment of Large Language Models (ICLR 2025 · similarity 0.88)
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.
