Kohsei Matsutani

I'm a 1st year master student at the University of Tokyo (where I also did my undergrad) under the supervision of Prof. Yutaka Matsuo. Before, I spent a year as an MTS intern at Liquid AI.
Now I’m working on (i) investigating whether LLMs/VLMs have implicit “world models” (causal understanding, abstraction, discovery, etc.) and what their limitations are, and (ii) improving post-training, especially RL, to ultimately make them continually adaptable in open-ended environments.

Email / Google Scholar / Linkedin / alphaXiv / Hugging Face

Research

	Few-Shot Demonstrations Elicit the Use of In-Context World Representations in LLMs Kohsei Matsutani, Gouki Minehishi, Core Francisco Park, Takeshi Kojima, Yusuke Iwasawa, Yutaka Matsuo, Preprint, 2026 arXiv / Code
	OrderGrad: Optimizing Beyond the Mean with Order-Statistic Policy Gradient Estimation Paavo Parmas, Yongmin Kim, Kohsei Matsutani, Shota Takashiro, Soichiro Nishimori, Takeshi Kojima, Yusuke Iwasawa, Yutaka Matsuo, Preprint, 2026 arXiv / Code
	On Advantage Estimates for Max@K Policy Gradients Shota Takashiro, Soichiro Nishimori, Paavo Parmas, Yongmin Kim, Kohsei Matsutani, Gouki Minegishi, Yusuke Iwasawa, Takeshi Kojima, Yutaka Matsuo, Preprint, 2026 arXiv
	Zipping the Thought: When and How Compressed Reasoning Data Works in LLM Post-Training Kohsei Matsutani, Gouki Minegishi, Takeshi Kojima, Yusuke Iwasawa, Yutaka Matsuo, EMNLP (Findings), 2026 arXiv / Code
	MultiBanana: A Challenging Benchmark for Multi-Reference Text-to-Image Generation Yuta Oshima, Daiki Miyake, Kohsei Matsutani, Yusuke Iwasawa, Masahiro Suzuki, Yutaka Matsuo, Hiroki Furuta CVPR, 2026 arXiv / Code / Hugging Face (100K+ DLs, 202610)
	RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs Kohsei Matsutani, Shota Takashiro, Gouki Minegishi, Takesi Kojima, Yusuke Iwasawa, Yutaka Matsuo ICLR, 2026 arXiv
	Mixture of Experts Provably Detect and Learn the Latent Cluster Structure in Gradient-Based Learning Ryotaro Kawata (), Kohsei Matsutani (), Yuri Kinoshita, Naoki Nishikawa, Taiji Suzuki ^* Equal contribution ICML, 2025 arXiv