Game Theory Driven Multi-Agent Framework Mitigates Language Model Hallucination

Large Language Models (LLMs) often struggle with scientific domains because they prioritize linguistic patterns over axiomatic reasoning, leading to frequent hallucinations. This study investigates whether a game-theoretic, multi-agent framework can force lightweight models to internalize physical constraints and structured reasoning, thereby improving their reliability in specialized fields like chemistry.

The authors developed G-Frame, an adaptive multi-agent architecture that operates on two levels. At the micro-level, a cooperative Team Game decomposes complex scientific tasks into manageable subtasks, reducing the entropy of generated text. At the macro-level, a Bayesian Game functions as a decisional agent that monitors the system state and updates policies under uncertainty. This framework was used to clean a 5-billion-token chemical corpus and synthesize over 500,000 chains-of-thought and question-answer pairs to fine-tune a 7B parameter model, OmniChem.

OmniChem demonstrates significant improvements in factual accuracy and reasoning capabilities compared to its base architecture. On the ChemJudge benchmark, the model achieved a 79.46% reduction in hallucinations. Furthermore, OmniChem shows performance parity with GPT-4o mini on standard chemistry benchmarks (ThChem and ChemBench) and exhibits advanced capabilities in molecular design, such as optimizing BODIPY derivatives for specific light absorption and water solubility, as well as generating accurate retrosynthetic plans for drugs like lidocaine.

This work provides a scalable paradigm for deploying specialized, lightweight LLMs in scientific research. By replacing open-ended generation with a structured, game-theoretic closed-loop, the framework offers a viable path to overcome the reasoning deficiencies that currently limit the utility of smaller models in high-stakes scientific discovery.