Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PersLLM: A Personified Training Approach for Large Language Models

Zheni Zeng, Jiayi Chen, Huimin Chen, Yukun Yan, Yuxuan Chen, Zhenghao Liu, Zhiyuan Liu, Maosong Sun

TL;DR

PersLLM presents a data-centric approach to LLM personification by integrating structured data construction (with anti-induction and CoT prompting, temporal staging, and objective/subjective material separation) and a two-step tuning pipeline (personified conversational tuning plus automatic DPO) to produce dynamic, distinctive personalities. The framework is validated theoretically via personality scale analyses and practically through HP-like and real-world case studies, showing improved personality distinction, consistency, and more natural human-agent interactions. Key contributions include detailed data construction strategies, an automated DPO-based tuning mechanism, and publicly available code and demonstrations. The results suggest PersLLM can enhance human-like interactions, multi-agent collaboration, and domain-specific personification, with implications for education, consultation, and social simulations.

Abstract

Large language models (LLMs) exhibit human-like intelligence, enabling them to simulate human behavior and support various applications that require both humanized communication and extensive knowledge reserves. Efforts are made to personify LLMs with special training data or hand-crafted prompts, while correspondingly faced with challenges such as insufficient data usage or rigid behavior patterns. Consequently, personified LLMs fail to capture personified knowledge or express persistent opinion. To fully unlock the potential of LLM personification, we propose PersLLM, a framework for better data construction and model tuning. For insufficient data usage, we incorporate strategies such as Chain-of-Thought prompting and anti-induction, improving the quality of data construction and capturing the personality experiences, knowledge, and thoughts more comprehensively. For rigid behavior patterns, we design the tuning process and introduce automated DPO to enhance the specificity and dynamism of the models' personalities, which leads to a more natural opinion communication. Both automated metrics and expert human evaluations demonstrate the effectiveness of our approach. Case studies in human-machine interactions and multi-agent systems further suggest potential application scenarios and future directions for LLM personification.

PersLLM: A Personified Training Approach for Large Language Models

TL;DR

PersLLM presents a data-centric approach to LLM personification by integrating structured data construction (with anti-induction and CoT prompting, temporal staging, and objective/subjective material separation) and a two-step tuning pipeline (personified conversational tuning plus automatic DPO) to produce dynamic, distinctive personalities. The framework is validated theoretically via personality scale analyses and practically through HP-like and real-world case studies, showing improved personality distinction, consistency, and more natural human-agent interactions. Key contributions include detailed data construction strategies, an automated DPO-based tuning mechanism, and publicly available code and demonstrations. The results suggest PersLLM can enhance human-like interactions, multi-agent collaboration, and domain-specific personification, with implications for education, consultation, and social simulations.

Abstract

Large language models (LLMs) exhibit human-like intelligence, enabling them to simulate human behavior and support various applications that require both humanized communication and extensive knowledge reserves. Efforts are made to personify LLMs with special training data or hand-crafted prompts, while correspondingly faced with challenges such as insufficient data usage or rigid behavior patterns. Consequently, personified LLMs fail to capture personified knowledge or express persistent opinion. To fully unlock the potential of LLM personification, we propose PersLLM, a framework for better data construction and model tuning. For insufficient data usage, we incorporate strategies such as Chain-of-Thought prompting and anti-induction, improving the quality of data construction and capturing the personality experiences, knowledge, and thoughts more comprehensively. For rigid behavior patterns, we design the tuning process and introduce automated DPO to enhance the specificity and dynamism of the models' personalities, which leads to a more natural opinion communication. Both automated metrics and expert human evaluations demonstrate the effectiveness of our approach. Case studies in human-machine interactions and multi-agent systems further suggest potential application scenarios and future directions for LLM personification.
Paper Structure (15 sections, 6 figures, 12 tables)

This paper contains 15 sections, 6 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Data Construction
  5. Model Tuning
  6. Experiment
  7. Corpus and Datasets
  8. Backbone Models and Baselines
  9. Theoretical Evaluation
  10. Practical Evaluation
  11. Discussion
  12. Hyper-parameter Settings
  13. Dataset Details
  14. Prompt Settings
  15. Interaction Cases

Figures (6)

  • Figure 1: Schematic diagram for PersLLM. (a) Collect raw data by category; (b) Conduct automatic annotation for conversational data; (c) Tune LLMs with personified data and automatic DPO; (d) Several data examples.
  • Figure 2: NEO-120 scale test results for Big Five personality traits analysis.
  • Figure 3: Case study on HP dataset. Comparison of : (a) personality attitudes; (b) knowledge reserves; (c) different time stages; (d) the personified and ordinary alignment methods on values; (e) personified and RoCIT models.
  • Figure 4: Human evaluation for Huiyin Lin agent.
  • Figure 5: Conflict instance for multi-agent communication. (a) PersLLM based on MiniCPM-2.4B; (b) GPT-4 with personality profile and prompt; (c) GPT-4 with only chatting history.
  • ...and 1 more figures