TRIO aligns three agents to refine retrieval-augmented generation
Original title: TRIO: Task Refinement with Individual Optimization for Retrieval-Augmented Generation
Query, judge, and answer agents are trained with separate rewards, including voting labels and curriculum matching.
The paper proposes TRIO, Task Refinement with Individual Optimization, for retrieval-augmented generation. Three agents iteratively refine queries, judge whether expanded queries are sufficient, and produce answers, each with its own optimization objective. A query agent is rewarded on improved answer quality, a judge is trained on unanimity-voting labels, and an answer agent uses dynamic matching rewards with curriculum learning. The authors report better results than state-of-the-art methods across multiple benchmarks and release code.
- A query extension agent is refined using reward from improved answer quality.
- Unanimity voting labels expanded queries as insufficient or sufficient for the judge agent.
- The answer agent uses dynamic matching-based rewards and curriculum learning.
- Code is released at the authors' GitHub repository.
From the original article
1 min read at huggingface.co
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.
By Yingtao Ren, Xiao Luo, Yiwei Fu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.