Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BaZi-Based Character Simulation Benchmark: Evaluating AI on Temporal and Persona Reasoning

Siyuan Zheng, Pai Liu, Xi Chen, Jizheng Dong, Sihan Jia

TL;DR

This work introduces BaZi-based persona reasoning for AI-driven character simulation, presenting the first QA dataset (Celebrity 50) and a BaZi-LLM system that fuses symbolic BaZi rule-mapping with large language models to generate temporally dynamic personas. The approach translates birth time/place/gender into structured symbolic features, interprets them with classical BaZi logic, and adapts prompts across health, career, wealth, relationships, and kinship domains. Empirical results on Celebrity 50 show substantial accuracy gains over baseline LLMs, with degradation when BaZi information is misaligned, underscoring the value of culturally grounded symbolic-LLM integration. The work highlights practical implications for scalable, realistic NPCs and narrative agents while outlining limitations and avenues for future improvements in domain knowledge, data diversity, and reasoning architectures.

Abstract

Human-like virtual characters are crucial for games, storytelling, and virtual reality, yet current methods rely heavily on annotated data or handcrafted persona prompts, making it difficult to scale up and generate realistic, contextually coherent personas. We create the first QA dataset for BaZi-based persona reasoning, where real human experiences categorized into wealth, health, kinship, career, and relationships are represented as life-event questions and answers. Furthermore, we propose the first BaZi-LLM system that integrates symbolic reasoning with large language models to generate temporally dynamic and fine-grained virtual personas. Compared with mainstream LLMs such as DeepSeek-v3 and GPT-5-mini, our method achieves a 30.3%-62.6% accuracy improvement. In addition, when incorrect BaZi information is used, our model's accuracy drops by 20%-45%, showing the potential of culturally grounded symbolic-LLM integration for realistic character simulation.

BaZi-Based Character Simulation Benchmark: Evaluating AI on Temporal and Persona Reasoning

TL;DR

This work introduces BaZi-based persona reasoning for AI-driven character simulation, presenting the first QA dataset (Celebrity 50) and a BaZi-LLM system that fuses symbolic BaZi rule-mapping with large language models to generate temporally dynamic personas. The approach translates birth time/place/gender into structured symbolic features, interprets them with classical BaZi logic, and adapts prompts across health, career, wealth, relationships, and kinship domains. Empirical results on Celebrity 50 show substantial accuracy gains over baseline LLMs, with degradation when BaZi information is misaligned, underscoring the value of culturally grounded symbolic-LLM integration. The work highlights practical implications for scalable, realistic NPCs and narrative agents while outlining limitations and avenues for future improvements in domain knowledge, data diversity, and reasoning architectures.

Abstract

Human-like virtual characters are crucial for games, storytelling, and virtual reality, yet current methods rely heavily on annotated data or handcrafted persona prompts, making it difficult to scale up and generate realistic, contextually coherent personas. We create the first QA dataset for BaZi-based persona reasoning, where real human experiences categorized into wealth, health, kinship, career, and relationships are represented as life-event questions and answers. Furthermore, we propose the first BaZi-LLM system that integrates symbolic reasoning with large language models to generate temporally dynamic and fine-grained virtual personas. Compared with mainstream LLMs such as DeepSeek-v3 and GPT-5-mini, our method achieves a 30.3%-62.6% accuracy improvement. In addition, when incorrect BaZi information is used, our model's accuracy drops by 20%-45%, showing the potential of culturally grounded symbolic-LLM integration for realistic character simulation.
Paper Structure (25 sections, 4 figures, 3 tables)

This paper contains 25 sections, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Datasets
  3. Method
  4. Overview
  5. BaZi Rule Mapping
  6. Interpretation via Classical Logic
  7. Scenario-Oriented Analysis
  8. Dynamic Persona Prompt Generation
  9. Experiments
  10. Evaluation Objectives
  11. Experimental Design
  12. Implementation Details
  13. Results and Analysis
  14. Case Study
  15. Differences in BaZi Theory Interpretation
  16. ...and 10 more sections

Figures (4)

  • Figure 1: Mainstream approaches to character simulation rely on either pretraining/finetuning on existing literary works or prompt-based conditioning. In contrast, we propose a novel Bazi-prompt framework, which encodes birth information (birthday, gender, and place of birth) into symbolic features. This framework enables finer-grained character simulation in terms of personality, temporal dynamics, and more diverse interactions with different environments.
  • Figure 2: Sample information input to LLM
  • Figure 3: Question and Birthplace Counts Across Countries
  • Figure 4: Our model is organized into four main components: (1) input layer for birth-related information (birthday, gender, place of birth), (2) BaZi rule analysis, (3) BaZi reasoning, and (4) scenario-specific interpretation. The BaZi-LLM prompt workflow outputs fine-grained features describing personality traits and dynamic states of daily interactions with external dimensions such as wealth, career, kinship, and health.