Researchers Propose Three-Agent RAG Framework TRIO

Researchers from the University of Technology Sydney proposed TRIO, a framework for retrieval-augmented generation (RAG), in a paper published on November 2. The method coordinates three specialized agents through iterative refinement, with separate optimization objectives for expanding queries, judging retrieved evidence and producing answers. The authors say current RAG systems can retrieve inadequate information and produce hallucinations.

TRIO's query expansion agent clarifies ambiguous questions and receives rewards based on whether the expanded query improves answer quality. A judge agent labels expanded queries as insufficient or sufficient through unanimous voting, then uses the resulting outputs as ground truth for its optimization. An answer agent is optimized with dynamic matching-based rewards and curriculum learning to improve evidence use and answer accuracy.

The authors report that TRIO achieved the highest average Exact Match across five question-answering benchmarks, outperforming strong RAG baselines including Search-R1. A Hugging Face summary reports 76.6% accuracy on unseen PubMedQA and says the approach also performed strongly with compact 4B and 8B models. The paper says experiments across multiple benchmark datasets validated the method against state-of-the-art approaches; its code is available on GitHub. The article provides no detailed benchmark results beyond these figures and comparisons.

Researchers from the University of Technology Sydney proposed TRIO, a framework for retrieval-augmented generation (RAG), in a paper published on November 2. The method coordinates three specialized agents through iterative refinement, with separate optimization objectives for expanding queries, judging retrieved evidence and producing answers. The authors say current RAG systems can retrieve inadequate information and produce hallucinations.

TRIO's query expansion agent clarifies ambiguous questions and receives rewards based on whether the expanded query improves answer quality. A judge agent labels expanded queries as insufficient or sufficient through unanimous voting, then uses the resulting outputs as ground truth for its optimization. An answer agent is optimized with dynamic matching-based rewards and curriculum learning to improve evidence use and answer accuracy.

The authors report that TRIO achieved the highest average Exact Match across five question-answering benchmarks, outperforming strong RAG baselines including Search-R1. A Hugging Face summary reports 76.6% accuracy on unseen PubMedQA and says the approach also performed strongly with compact 4B and 8B models. The paper says experiments across multiple benchmark datasets validated the method against state-of-the-art approaches; its code is available on GitHub. The article provides no detailed benchmark results beyond these figures and comparisons.