TY - JOUR
T1 - Enhancing belief consistency of large language model agents in decision-making process based on attribution theory
AU - He, Guoxiu
AU - Zhang, Meicong
AU - Su, Tiancheng
AU - Ma, Li
AU - Zhu, Xiaomin
N1 - Publisher Copyright:
© 2025 Elsevier Ltd
PY - 2026/2/1
Y1 - 2026/2/1
N2 - In social simulation, maintaining consistent beliefs (e.g., core cognition, position, and goal) is crucial for Large Language Model (LLM) agents to effectively assume specific roles and make coherent decisions. However, LLMs generate responses incrementally through next-token predictions based on preceding words, making them vulnerable to context ambiguity. This can cause their decision-making behavior to deviate from initial goals. To address this issue, we propose the Attribution-based Decision-Making Approach (ADMA), inspired by human attribution processes. ADMA comprises three modules: generation, evaluation, and attribution. Rather than making impulsive decisions, ADMA encourages LLMs to systematically attribute the underlying beliefs associated with each potential choice, and then adjust their decisions accordingly. Through iterative application, ADMA promotes better alignment between generated decisions and predefined goals. We evaluate ADMA in a debate scenario, where beliefs are continuously challenged by opposing arguments, influencing both stance and rebuttal strength. Empirical results show that the attribution process enhances belief consistency, as well as fluency, persuasiveness, and confidence. ADMA also improves LLM agents’ resistance to persuasive language. Furthermore, belief consistency increases with more attribution iterations, albeit with diminishing returns. These findings underscore the potential of attribution theory to strengthen belief stability in LLM agents.
AB - In social simulation, maintaining consistent beliefs (e.g., core cognition, position, and goal) is crucial for Large Language Model (LLM) agents to effectively assume specific roles and make coherent decisions. However, LLMs generate responses incrementally through next-token predictions based on preceding words, making them vulnerable to context ambiguity. This can cause their decision-making behavior to deviate from initial goals. To address this issue, we propose the Attribution-based Decision-Making Approach (ADMA), inspired by human attribution processes. ADMA comprises three modules: generation, evaluation, and attribution. Rather than making impulsive decisions, ADMA encourages LLMs to systematically attribute the underlying beliefs associated with each potential choice, and then adjust their decisions accordingly. Through iterative application, ADMA promotes better alignment between generated decisions and predefined goals. We evaluate ADMA in a debate scenario, where beliefs are continuously challenged by opposing arguments, influencing both stance and rebuttal strength. Empirical results show that the attribution process enhances belief consistency, as well as fluency, persuasiveness, and confidence. ADMA also improves LLM agents’ resistance to persuasive language. Furthermore, belief consistency increases with more attribution iterations, albeit with diminishing returns. These findings underscore the potential of attribution theory to strengthen belief stability in LLM agents.
KW - Agent
KW - Attribution theory
KW - Belief consistency
KW - Decision-making
KW - Large language model
UR - https://www.scopus.com/pages/publications/105013147781
U2 - 10.1016/j.eswa.2025.129273
DO - 10.1016/j.eswa.2025.129273
M3 - 文章
AN - SCOPUS:105013147781
SN - 0957-4174
VL - 297
JO - Expert Systems with Applications
JF - Expert Systems with Applications
M1 - 129273
ER -