Mengyu Ye
Ph.D. student @ Tohoku University Fundamental AI lab.
ye.mengyu.s1 [at] dc.tohoku.ac.jp
I am a final-year Ph.D. student in NLP and machine learning at Tohoku University, advised by Prof. Jun Suzuki, and a Google PhD Fellow. This autumn I will join Prof. Mrinmaya Sachan’s group at ETH Zürich as a visiting researcher.
My research uses controlled experiments to find where LLMs break and why, from reasoning failures to the limits of the tools we use to interpret them. The same approach keeps producing new results at NeurIPS, ACL, and EMNLP.
Most recently, I led Sumi, a fully open 7B diffusion language model pretrained from scratch on 1.5T tokens and matching autoregressive baselines, released for the community to build on.
news
| Jun 23, 2026 | We release Sumi, a fully open 7B uniform diffusion language model pretrained from scratch on 1.5T tokens, with weights, checkpoints, and the full training recipe — paper now on arXiv. |
|---|---|
| Jun 22, 2026 | I will join Prof. Mrinmaya Sachan’s group at ETH Zürich as a visiting researcher from autumn 2026 to March 2027. |
| Jan 30, 2026 | We release a new paper on relaxing positional alignment in masked diffusion LMs, identifying a key failure mode in open-ended generation, now on arXiv. |
| Dec 08, 2025 | Our team won the Best Static Evaluation Prize in the MMU-RAG NeurIPS 2025 Competition. |
| Dec 01, 2025 | Released a CLI tool that uses an LLM agent to automatically clean, format, and update BibTeX references. |