Asahi Shimbun Research Paper Accepted at International Conference 'ACL 2026'

The Asahi Shimbun Company announced that a research paper authored by members of its Media Research and Development Center has been accepted for presentation at the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), a premier international conference in the field of natural language processing.

The paper, with Akira Kawabata of the Media Research and Development Center as the lead author, introduces 'Cooperative yet Critical reward modeling (C2),' a novel method designed to more accurately evaluate the quality of answers generated by Large Language Models (LLMs).

As the use of generative AI expands, effectively evaluating the quality of AI-generated responses has become a critical challenge. This research proposes a new framework using rubrics to stabilize and refine evaluation criteria, which is expected to contribute to improving the reliability of generative AI.

[Overview of the C2 Framework] To address the challenges in current evaluation methods, the research team introduces two AIs with different roles: an AI that proposes rubrics and an AI that uses those rubrics to judge the quality of answers. By iteratively generating, evaluating, and learning from multiple rubrics, the system automatically collects pairwise data that guides both models toward more accurate decision-making.

[Results] Experimental results show that the C2 method outperforms existing approaches in evaluation accuracy. Furthermore, it was confirmed that models trained using the C2 framework achieved performance levels comparable to those using rubrics generated by models four times their size.

The research results will be presented at ACL 2026, which will be held in San Diego, U.S.A., from July 2 to July 7, 2026.