Presentation Information
[4Yin-A-58]Assigning personality traits to reasoning models via their thought processes
〇Tetsuya Hashimoto1, Mitsuhiro Mabuchi1, Miki Ota2, Taisei Ozaki2 (1. TOYOTA MOTOR CORPORATION, 2. MATSUO INSTITUTE, INC)
Keywords:
LLM,personality traits,machine learning
The construction of trust between humans and AI is generally considered to consist of two components: affective trust, which is based on humane, and cognitive trust, which is based on performance factors such as task execution accuracy.
This study aims to realize affective trust in human–AI relationships. As a first step, we seek to develop an AI that possesses human-like reasoning processes and personality traits. To enable explicit learning of reasoning and personality, we focus on a reasoning large language model (Reasoning model) that can output its thought processes, and attempt to attach personality traits to it.
While previous studies have explored the assignment of personality traits to conventional LLMs, research targeting Reasoning models remains limited. Therefore, we conducted a comparative evaluation of personality trait acquisition in Reasoning models using three representative training methods: In-context Learning (ICL), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO).
We compared these methods in terms of the degree of personality expression and general task performance accuracy. The results demonstrate the advantages and challenges of GRPO relative to the other methods, and we discuss directions for future research.
This study aims to realize affective trust in human–AI relationships. As a first step, we seek to develop an AI that possesses human-like reasoning processes and personality traits. To enable explicit learning of reasoning and personality, we focus on a reasoning large language model (Reasoning model) that can output its thought processes, and attempt to attach personality traits to it.
While previous studies have explored the assignment of personality traits to conventional LLMs, research targeting Reasoning models remains limited. Therefore, we conducted a comparative evaluation of personality trait acquisition in Reasoning models using three representative training methods: In-context Learning (ICL), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO).
We compared these methods in terms of the degree of personality expression and general task performance accuracy. The results demonstrate the advantages and challenges of GRPO relative to the other methods, and we discuss directions for future research.