TRIO: Task Refinement with Individual Optimization for Retrieval-Augmented Generation
Abstract
TRIO coordinates three specialized agents through iterative refinement and individual optimization objectives to improve retrieval-augmented generation by enhancing query expansion, retrieval judgment, and answer accuracy.
This paper studies retrieval-augmented generation (RAG), which leverages external knowledge to improve the performance of language models. Despite the remarkable progress, current RAG approaches are still far from satisfactory due to inadequate retrieval and potential hallucinations. Towards this end, we propose a novel approach named Task Refinement with Individual Optimization (TRIO) for RAG. The core idea of our TRIO is to equip three well-designed agents with an iterative refinement paradigm by aligning their individual optimization objectives. Specifically, we first expand ambiguous queries with a query extension agent and quantify the reward based on the improved answer quality, which can be utilized to refine our agent for higher expansion efficiency. To enhance the retrieval quality, we include a unanimity voting to annotate the expanded query as insufficient or sufficient, and their generative outcomes are utilized as ground truth to optimize the judge agent. To further improve evidence utilization and answer accuracy, an answer agent is optimized with dynamic matching-based rewards with curriculum learning for final outputs. Extensive experiments across multiple benchmark datasets validate the effectiveness of the proposed TRIO in comparison with state-of-the-art approaches. Our code is available at https://github.com/yingtaoren/TRIO.
Community
Can three specialized agents make RAG more accurate than a single powerful reasoning model? 🚀
Introducing TRIO, a multi-agent RAG framework that improves retrieval-augmented reasoning through specialized agent optimization.
TRIO jointly coordinates three agents, each optimized for a distinct role:
Query Agent: Learns to retrieve evidence that directly improves downstream answer quality.
Judge Agent: Determines whether the retrieved evidence is sufficient to answer the question.
Answer Agent: Develops robust reasoning capabilities over noisy retrieved contexts through reward-driven curriculum learning.
Key highlights:
- Top average Exact Match across five QA benchmarks, outperforming strong RAG baselines, including Search-R1.
- 76.6% accuracy on unseen PubMedQA, demonstrating promising cross-domain generalization.
- Strong performance with compact 4B and 8B models, highlighting the potential of specialized multi-agent optimization.
Key insight: Effective RAG depends on how well retrieval, evidence assessment, and answer generation work together. Optimizing each agent for its specific role can substantially improve end-to-end reasoning.
Get this paper in your agent:
hf papers read 2611.02810 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper