Rikiya Takehi

I am a first-year PhD student at MIT CSAIL, advised by Prof. Omar Khattab. I recently received a B.Eng. from Waseda University, where I was supervised by Prof. Tetsuya Sakai. Previously, I was a research intern at NVIDIA Research and Mixedbread. Before that, I was a guest researcher at NIST working with Dr. Ian Soboroff and Dr. Ellen Voorhees, and I have also collaborated with Prof. Fernando Diaz of CMU LTI on retrieval rankings.

My graduate studies are supported by the Toyota PhD Fellowship and the MIT Distinguished Science and Engineering Fellowship.

I am interested in Machine Learning (ML), Natural Language Processing (NLP), and Information Retrieval (IR). Now, I am especially interested in designing better LLMs / LLM Systems / LLM Agents.

News

Publications

You can also find my articles on my Google Scholar profile.

1. MLLMs Fail to Refuse when Using Tools Agentically
   Rikiya Takehi, Ryo Hachiuma, Shaona Ghosh, Dan Zhao, Frank Wang, Yusuke Hirota. 2026.
   NeurIPS 2026 Eval & Data Track (acceptance rate: 26%).
   arXiv
2. Diversification as Risk Minimization
   Rikiya Takehi, Fernando Diaz, Tetsuya Sakai. 2025.
   WSDM 2026.
   Best Paper Award (top 0.1%, 1 of 799 submissions)
   arXiv
3. Fantastic (small) Retrievers and How to Train Them: mxbai-edge-colbert-v0 Tech Report.
   Rikiya Takehi, Benjamin Clavié, Sean Lee, Aamir Shakir. 2025.
   Tech Report.
   3M+ downloads on Hugging Face
   Tech Report | Blog | 17M ColBERT model | 32M ColBERT model
4. Beyond Match Maximization and Fairness: Retention-Objectified Two-Sided Matching
   Ren Kishimoto^*, Rikiya Takehi^*, Koichi Tanaka, Masahiro Nomura, Riku Togashi, Yuta Saito. 2025.
   ICLR 2026.
   preprint
5. General Framework for Off-Policy Learning with Partially-Observed Reward
   Rikiya Takehi, Kosuke Kawakami, Masahiro Asami, Yuta Saito. 2025.
   ICLR 2025.
   arXiv | OpenReview | presentation | poster
6. LLM-Assisted Relevance Assessments: When Should We Ask LLMs for Help?
   Rikiya Takehi, Ellen M. Voorhees, Tetsuya Sakai, and Ian Soboroff. 2025.
   SIGIR 2025.
   arXiv | slides | code
7. Open-Source LLM-based Relevance Assessment vs. Highly Reliable Manual Relevance Assessment: A Case Study
   Tetsuya Sakai, Khant Myoe Rain, Rikiya Takehi, Sijie Tao, Youngin Song. 2025.
   CIKM 2025.
   proceedings
8. Objective-driven Calibrated Recommendations
   Rikiya Takehi^*, Koichi Tanaka^*, Ren Kishimoto, Masahiro Nomura, Riku Togashi, Yuta Saito. 2025.
   preprint
9. Open-Domain Dialogue Quality Evaluation: Deriving Nugget-level Scores from Turn-level Scores
   Rikiya Takehi, Akihisa Watanabe, and Tetsuya Sakai. 2023.
   SIGIR-AP 2023.
   code | poster | slides | proceedings

Experience

• Mar 2026 - Aug 2026 · remote
  Research Intern, NVIDIA Research
• Aug 2025 - Feb 2026 · CA, USA
  Research Intern, Mixedbread
• Oct 2024 - Sep 2025 · Tokyo, Japan
  Research Intern, CyberAgent AI Lab. Algorithm Team
• Sept 2024 - Apr 2025 · Tokyo, Japan
  Research Intern, Hakuhodo Tech Inc.
• Oct 2023 - Sep 2024 · Maryland, US
  Guest Researcher, NIST Retrieval Group

Achievements

Selected Invited Talks

• Feb 2026 · Invited by Dr. Thorsten Joachims (Cornell)
  Cornell University
  Diversification as Risk Minimization
• Jul 2025 · Invited by Dr. Tracy Holloway King (Adobe)
  SIGIR eCom Workshop 2025
  Product Search and Recommendations (with Prof. ChengXiang Zhai)
• Jun 2025 · Invited by Prof. Charles Clarke (UWaterloo)
  EVIA 2025
  Using LLMs as Assistants for Building Large Test Collections
• Aug 2024 · Invited by Prof. Douglas Oard (UMD)
  University of Maryland, College Park
  Nugget-Based Evaluation and the Use of LLMs
• 2025 · Panel
  NTCIR 2025 Panelist
  With Profs. Maarten de Rijke (UvA), Mark Sanderson (RMIT), Charles Clarke (UWaterloo), and Ian Soboroff (NIST).