Research map / Large language models / Language, safety and interpretability
The Language Model Can Have the Personality: Joint Learning for Personality Enhanced Language Model (Student Abstract)
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 PELM, a joint learning framework that trains a GPT-2/BERT-based MBTI personality classifier and a language model with shared weights and a summed loss, so generated text reflects personality types.
Problem
Chatbots lack relatable, human-like personality; existing persona approaches inject profiles or background info, which is insufficient to make language models personality-consistent.
Approach
Jointly train a transformer classifier (MBTI type prediction) and a transformer decoder language model with parameter sharing across all layers and a joint loss L_lm + γ·L_cls, so personality information is injected at training time via both labels and text.
Key results
- PELM with GPT-2 raises classification weighted F1 from 59.56% (clsGPT-2) to 61.05%.
- PELM with GPT-2 improves P-Index from 47.4 to 50.0 versus lmGPT-2, while perplexity rises slightly (52.28 to 52.77).
- PELM with BERT improves F1 from 60.19% to 60.68% and P-Index from 17.7 to 19.9.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use joint classification and language modeling instead of classification-only training | BERT · MBTI | weighted F1 (%): 60.19 -> 60.68 | Helped |
| Use joint classification and language modeling instead of classification-only training | GPT-2 · MBTI | weighted F1 (%): 59.56 -> 61.05 | Helped |
| Use joint classification and language modeling instead of language-model-only training | BERT · MBTI | P-Index: 17.7±3.9 -> 19.9±4.0 | Helped |
| Use joint classification and language modeling instead of language-model-only training | GPT-2 · MBTI | P-Index: 47.4±2.6 -> 50.0±3.1 | Helped |
| Use joint classification and language modeling instead of language-model-only training | BERT · MBTI | Perplexity: 75.63 -> 77.31 | Hurt |
| Use joint classification and language modeling instead of language-model-only training | GPT-2 · MBTI | Perplexity: 52.28 -> 52.77 | Hurt |
Surprising findings
- Joint training gives a slightly higher perplexity for the GPT-2 language model (52.28 to 52.77), so the personality gain comes with a small fluency cost.
Limitations
- Evaluation is small-scale: a single dataset, a preliminary study, and a single case study with one prompt.
- Perplexity does not improve for the GPT-2 setting, and the P-Index metric is a keyword-count proxy for personality.
- No human evaluation or LLM-based judging of personality expression.
Experimental setup
Datasets: MBTI dataset (Kaggle, datasnaek/mbti-type)
Benchmarks: MBTI dataset
Metrics: weighted F1 perplexity P-Index (TF-IDF keyword counts)
Baselines: clsBERT lmBERT clsGPT-2 lmGPT-2
Builds on: BERT GPT-2
Closest papers by meaning
- Neuron based Personality Trait Induction in Large Language Models (ICLR 2025 · similarity 0.86)
- PICLe: Eliciting Diverse Behaviors from Large Language Models with Persona In-Context Learning (ICML 2024 · similarity 0.85)
- Open Models, Closed Minds? On Agents Capabilities in Mimicking Human Personalities through Open Large Language Models (AAAI 2025 · similarity 0.85)
- Personality Alignment of Large Language Models (ICLR 2025 · similarity 0.84)
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models (NeurIPS 2026 · similarity 0.84)
- LLM vs Small Model? Large Language Model Based Text Augmentation Enhanced Personality Detection Model (AAAI 2024 · similarity 0.84)
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.