Game theory driven multi-agent framework mitigates language model hallucination
Abstract
Lightweight large language models hold promise for scientific applications but remain limited in rule-constrained domains because they can reproduce linguistic patterns without following scientific principles, leading to hallucinations. Here, we show that G-Frame, an adaptive multi-agent framework inspired by Bayesian game and team game principles, enables a closed-loop process for automated data generation, model evaluation, and training. By assigning complementary roles to multiple agents, G-Frame generates structured reasoning data and constructs a specialized corpus containing 363,045 chain-of-thought examples and 199,589 question–answer pairs. Using this corpus together with domain pre-training and adaptive training, the 7-billion-parameter model OmniChem achieves competitive performance across the reported benchmarks and reduces the ChemJudge hallucination rate from 41.95% for the base model to 12.91%. Case studies further illustrate potential applications in molecular design and selected synthesis-planning tasks. These results support G-Frame as an extensible game theory-driven multi-agent framework for improving the reliability of lightweight scientific language models.
Related
- A unified multi-domain framework for hallucination detection and reliability evaluation in large language models
- Large Language Models in Chemistry: Opportunities, Limitations, and Hallucination Risks
- BrainAgent: A Large Language Model-Driven Multi-Agent Framework for Autonomous Brain Signal Understanding
- Hybrid retrieval–verification framework for hallucination reduction in medical LLMs
- Hallucination in Time-series Large Language Models: An empirical lnvestigation and analysis of mitigation strategies
BibTeX
@article{liu2026game,
title = {{Game theory driven multi-agent framework mitigates language model hallucination}},
author = {Runzhe Liu and Biquan Bie and Zihao Wang and Yuchao Ma and Yexin Liu and Xinghai Li and Harry Yang and Wenbo Yang and Jinzhe Cao and Shengyang Tao},
journal = {Nature Communications},
year = {2026},
month = oct,
doi = {10.1038/s41467-026-78213-5},
url = {https://www.nature.com/articles/s41467-026-78213-5},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - Game theory driven multi-agent framework mitigates language model hallucination
AU - Runzhe Liu
AU - Biquan Bie
AU - Zihao Wang
AU - Yuchao Ma
AU - Yexin Liu
AU - Xinghai Li
AU - Harry Yang
AU - Wenbo Yang
AU - Jinzhe Cao
AU - Shengyang Tao
PY - 2026
DA - 2026/10/07
JO - Nature Communications
DO - 10.1038/s41467-026-78213-5
UR - https://www.nature.com/articles/s41467-026-78213-5
AB - Lightweight large language models hold promise for scientific applications but remain limited in rule-constrained domains because they can reproduce linguistic patterns without following scientific principles, leading to hallucinations. Here, we show that G-Frame, an adaptive multi-agent framework inspired by Bayesian game and team game principles, enables a closed-loop process for automated data generation, model evaluation, and training. By assigning complementary roles to multiple agents, G-Frame generates structured reasoning data and constructs a specialized corpus containing 363,045 chain-of-thought examples and 199,589 question–answer pairs. Using this corpus together with domain pre-training and adaptive training, the 7-billion-parameter model OmniChem achieves competitive performance across the reported benchmarks and reduces the ChemJudge hallucination rate from 41.95% for the base model to 12.91%. Case studies further illustrate potential applications in molecular design and selected synthesis-planning tasks. These results support G-Frame as an extensible game theory-driven multi-agent framework for improving the reliability of lightweight scientific language models.
SN - 2041-1723
ER -