Multi-Agent Framework Reduces LLM Hallucinations in Science

Lightweight Large Language Models (LLMs) often struggle in scientific domains governed by strict rules, frequently generating "hallucinations" by mimicking linguistic patterns instead of applying axiomatic reasoning. This limitation severely restricts their practical application in fields requiring high accuracy. Researchers have introduced G-Frame, an adaptive multi-agent framework designed to mitigate these hallucinations. G-Frame uses Bayesian and team game principles to create an automated, closed-loop system for synthesizing high-quality data and training models. By forcing LLMs to internalize domain constraints through structured reasoning, the framework generated a specialized corpus of chains-of-thought and question-answer pairs. This process led to the development of OmniChem, a 7B parameter model that achieves performance parity with GPT-4o mini on specific benchmarks while demonstrating a 79.46% reduction in hallucinations compared to its base architecture. OmniChem also shows advanced capabilities in molecular design and synthesis planning, establishing a scalable paradigm for accelerating knowledge discovery in specialized scientific fields.