Akira Kawabata
Ph.D. student at SOKENDAI, affiliated with the National Institute of Informatics (NII). Advised by Saku Sugawara.
Research interests Scalable Oversight, Alignment
Publications
-
C2: Scalable Rubric-Augmented Reward Modeling from Binary Preferences
ACL 2026 · Oral
We show that rubric planning and verification behaviors can be synthesized from simple binary preferences. The resulting rubric-augmented reward models align more closely with human preferences.
-
Rationale-Aware Answer Verification by Pairwise Self-Evaluation
EMNLP 2024
We show that verifiers trained on outcome-based labels fail to distinguish sound reasoning from flawed reasoning that happens to reach the correct answer, and that refining the training data through the model’s own self-evaluation yields verifiers that favor well-reasoned solutions.
-
Evaluating the Rationale Understanding of Critical Reasoning in Logical Reading Comprehension
EMNLP 2023
We build a benchmark, based on multiple-choice logical reasoning questions, that tests whether models can evaluate the rationales behind answers. The gap to humans is largest in judging why the incorrect answers are wrong, which suggests that choosing the right answer does not imply understanding the reasoning behind it.
News
- Apr 2026 Our paper on scalable rubric-augmented reward modeling is accepted to ACL 2026.
- Mar 2025 Received the INMA 30 Under 30 Award.
- Sep 2024 Our paper on LLM-based rationale verification is accepted to EMNLP 2024.
Experience
- Apr 2023 — Mar 2026 Research Engineer, The Asahi Shimbun Company
Education
- Apr 2026 — present Ph.D. student, SOKENDAI / NIIDepartment of Informatics. Supervisor: Saku Sugawara.
- Apr 2021 — Mar 2023 M.Eng., Nara Institute of Science and TechnologySupervisor: Taro Watanabe.
- Apr 2016 — Mar 2021 B.A., Hitotsubashi University
Honors and scholarships
- 2026 SOKENDAI Special Researcher Program, supported by JST BOOST.
- 2025 International News Media Association (INMA) 30 Under 30 Awards.
- 2023 Young Researcher's Encouragement Award, 29th Annual Meeting of the Association of Natural Language Processing.