Researchers from Tsinghua and South China University of Technology developed the Atomic Task Graph (ATG) framework, allowing 7B-8B open-source models to beat GPT-4 on two out of three complex agent benchmarks without fine-tuning or extra training. ATG decomposes complex tasks into directed graphs of atomic tool calls, enabling parallel execution and localized failure repair, significantly reducing hallucination rates to 12% compared to ReAct's 43%. The ATG framework achieved a score of 63.65 on ALFWorld using an 8B Llama model, outperforming GPT-4 with ReAct which scored 41.24 on the same benchmark.
A new agent architecture significantly improves small model performance on complex tasks, directly reducing the compute required for agentic AI workloads.