On 10th March 2026, Sho Fukada (B4 at the time of award), Department of Electrical Engineering and Information Systems , won 1st Place in the Retrieval Performance Category and 3rd Place in the RAG Category at TREC 2025 RAG Track.
TREC 2025 RAG Track, 1st Place in the Retrieval Performance Category and 3rd Place in the RAG Category
The University of Tokyo & Hitotsubashi University Team Achieves Top Rankings at the International Information Retrieval Evaluation Workshop "TREC 2025 RAG Track" (1st Place in the Retrieval Performance Category and 3rd Place in the RAG Category)
About awarded research
The joint team "UTokyo-HitU" from The University of Tokyo and Hitotsubashi University (Sho Fukada, Department of Electrical Engineering and Information Systems, School of Engineering, The University of Tokyo; Lecturer Yusuke Matsui, Graduate School of Information Science and Technology, The University of Tokyo; and Associate Professor Atsushi Keyaki, Faculty/Graduate School of Social Data Science, Hitotsubashi University) participated in the Retrieval-Augmented Generation Track (TREC RAG Track) of the international information retrieval evaluation workshop "TREC 2025," organized primarily by the U.S. National Institute of Standards and Technology (NIST). The team placed 1st in the Retrieval Performance category and 3rd in the RAG category.
TREC is an international evaluation workshop that compares and validates the effectiveness of information retrieval technologies using shared datasets and evaluation metrics. In the TREC RAG Track, the performance of Retrieval-Augmented Generation (RAG) systems is evaluated, where large language models generate answers based on evidence retrieved from large-scale document collections. In the 2025 edition of the track, participants were challenged to accurately retrieve relevant documents and generate evidence-based answers to complex information requests spanning multiple sentences. Approximately 20 teams participated in the track in 2025.
The UTokyo-HitU team developed a RAG pipeline that integrates retrieval, pseudo-answer generation, re-ranking, and answer generation, aiming to accurately retrieve evidence documents useful for answer generation in response to complex natural-language information requests. As a result, the team ranked 1st among participating systems in the Retrieval Performance category and 3rd in the RAG category, demonstrating the effectiveness of integrating retrieval performance with pseudo-answer generation in system design.
RAG is a technology that aims to provide more accurate and verifiable information by grounding generative AI responses in external documents. This achievement is expected to contribute to a wide range of applications that require evidence-based information presentation, including academic literature searches, private document retrieval, educational support, government, healthcare, and legal services.
TREC RAG Track Official Website: https://trec-rag.github.io/
Results Overview: https://arxiv.org/abs/2603.09891
Method Report by the UTokyo-HitU Team:
https://trec.nist.gov/pubs/trec34/papers/UTokyo.rag.pdf
Full List of Reports: https://pages.nist.gov/trec-browser/trec34/rag/proceedings/
Your impression & future plan
In RAG systems, it is important not only to generate natural-sounding text but also to accurately retrieve the documents that serve as evidence for the answers and effectively connect that evidence to the generated responses. We are very pleased that the UTokyo-HitU team's system received high recognition at this international evaluation workshop. In the future, we will continue our research toward developing retrieval and generative AI technologies that are more reliable, more practical for real-world applications, and faster.
Ayato Kitadai (M1), Department of Technology Management for Innovation, received Behavioral Economics Association Poster Presentation uragement Award (General Category)
