Abstract
The application of lightweight Large Language Models in rule-based scientific domains is severely limited by their tendency to mimic linguistic patterns rather than reproduce axiomatic reasoning, leading to frequent hallucinations. Here, we introduce G-Frame, an adaptive multi-agent framework integrating Bayesian and team game principles, establishing an automated closed-loop for high-quality data synthesis and model training.
By enforcing the internalization of domain constraints through structured reasoning, we synthesized a specialized corpus of 363,045 chains-of-thought and 199,589 question-answer pairs.
The resulting 7B model OmniChem achieves performance parity with GPT 4o mini on custom benchmarks and ChemBench, while exhibiting a 79.46% reduction in hallucinations relative to its base architecture. We further demonstrate the advanced capabilities of OmniChem in molecular design and synthesis planning.
This work establishes a scalable paradigm utilizing adaptive multi-agents to overcome inherent reasoning deficiencies, offering a feasible pathway for accelerating knowledge discovery in specialized scientific fields.
Blogger's Review: This study successfully reduces hallucinations in language models by incorporating game theory and multi-agent frameworks, showcasing significant potential for applications in scientific domains. The methodology of G-Frame offers new insights for future research, especially in data synthesis and model training, warranting further exploration.