Research map / Large language models / Alignment and preferences
Controllable and Explainable Personality Sliders for LLMs at Inference Time
Florian Hoppe, David Khachaturov, Robert D. Mullins, Mark Huasong Meng
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 |
|---|---|---|---|
| train subsequent probes on activations shifted by prior probes (Sequential Adaptive Steering) instead of independent probes (ablation) | multi-trait activation steering · E+ N+ A- profile | perplexity: 27.16 -> 14.54 | Improved |
| train subsequent probes on shifted activations (ablation) | multi-trait steering · E+ N+ A- profile | Extraversion score: 3.62 -> 4.20 | Improved |
| train subsequent probes on shifted activations (ablation) | multi-trait steering · E+ N+ A- profile | Neuroticism score: 3.10 -> 4.1 | Improved |
| add Fisher-ratio automated layer selection instead of random layers (ablation) | SAS · E+ N+ A- profile | Extraversion score: 3.30 -> 4.20 | Improved |
| add Fisher-ratio layer selection instead of random layers (ablation) | SAS · E+ N+ A- profile | perplexity: 14.79 -> 14.54 | No clear effect |
| replace independent probes with Gram-Schmidt orthogonalization post hoc | naive steering · perplexity at alpha 1.0 | perplexity: 17.10 -> 17.15 | No clear effect |
| use SAS instead of naive or Gram-Schmidt at alpha 1.5 | multi-trait steering · perplexity at alpha 1.5 | perplexity: > 83 -> 18.65 | Improved |
| use SAS instead of LoRA Soups merging three adapters | multi-trait personality control · E+ N+ A- profile | trait shift: Extraversion -0.12 -> +1.00 | Improved |
| use SAS instead of DPO fine-tuned model | multi-trait personality control · E+ N+ A- profile | trait shift | Improved |
| add SAS steering | Llama-3-8B-Instruct · TruthfulQA and HellaSwag | accuracy: 0.32 -> 0.32; 0.64 -> 0.64 | No clear effect |
| add SAS steering | Llama-3-8B-Instruct · generation quality | perplexity: 9.95 -> 14.54 | Hurt |
| orthogonalize probes with SAS training | personality steering vectors · inter-probe cosine similarity | off-diagonal cosine similarity | Improved |
Closest papers by meaning
- Personality Alignment of Large Language Models (ICLR 2025 · similarity 0.88)
- Neuron based Personality Trait Induction in Large Language Models (ICLR 2025 · similarity 0.88)
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models (NeurIPS 2026 · similarity 0.87)
- Open Models, Closed Minds? On Agents Capabilities in Mimicking Human Personalities through Open Large Language Models (AAAI 2025 · similarity 0.86)
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.86)
- PICLe: Eliciting Diverse Behaviors from Large Language Models with Persona In-Context Learning (ICML 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.
