News

Oct 21, 2026Research

A paper on AI agents by Harsh Chandra has been accepted at IEEE Big Data

Automatically translated from the Japanese original.

Paper Accepted

"MARC-RAG: Multi-Agent Routing and Co-adaptation for RAG," a new paper on AI agents authored by mign members and technical advisors, has been accepted to the IEEE International Conference on Big Data (IEEE Big Data 2026), one of the world's top conferences in the field.

About IEEE Big Data

IEEE Big Data is a world-leading international conference in the big data field, organized by IEEE, the world's largest professional society for electrical and information engineering. Papers submitted by researchers and companies from around the globe are selected through rigorous peer review, and each year the conference showcases state-of-the-art results spanning everything from fundamental big data research to industrial applications. IEEE Big Data 2026 will be held in Phoenix, Arizona, USA, from December 14 to 17, 2026.

Visit the official IEEE Big Data 2026 website

Overview of the Accepted Paper

The accepted paper advances RAG (Retrieval-Augmented Generation), the technique of combining document retrieval with language-model generation, by having multiple AI agents work together.

RAG is evolving from fixed retrieval pipelines into agentic systems that autonomously plan, retrieve, and generate answers. Existing approaches, however, face a trade-off: if the system configuration is fixed, the whole pipeline can be trained end to end, but the configuration cannot be adapted to each query; in dynamic approaches, only the Planner, which handles planning, is trained, while the Executors that carry out the work remain fixed. MARC-RAG, the dynamic Agentic RAG framework proposed in this paper, builds the Planner and a set of specialized Executors on a shared language model and optimizes them all simultaneously, using both task performance and efficiency as rewards, thereby achieving co-adaptation between the two. In experiments on seven open-domain question-answering benchmarks, MARC-RAG achieved an average F1 score of 53.86, outperforming the strongest baseline by 8.29 points on average.

Authors

Harsh Chandra (Ph.D., mign)
Masahiro Taima (Ph.D., mign)
Jikai Li (Ph.D., mign)
Hiroki Yoda (mign)
Bui Quang Thanh (mign)
Daisuke Kawahara (Ph.D., Waseda University, Kawahara Laboratory)
Yasushi Ikeda (Ph.D., The University of Tokyo, Ikeda Laboratory)

mign will continue to advance research and development on AI agents in collaboration with universities and research institutions, and to bring the results to society in the form of real products.