αlphabell.
α · People · Tomoko Niwa
Stylised avatar of Tomoko Niwa

Tomoko Niwa

Long-tenured contributor

Based in Kyoto
Node-cell voronoi-19
ORCID 0000-5560-3816-8152

Research

Tomoko's work sits between mathematics and machine learning: she develops the formal machinery behind operator-based latent dynamics — the conditions under which learned operators compose, the bounds on what they can extrapolate to, and the geometric structure that makes some operator libraries reusable across modalities.

She is a co-author on the compositional latent dynamics result and is currently leading a multi-modal extension that targets unified perception-prediction at scale. She presents results at NeurIPS and ICML with what one collaborator calls 'a complete absence of marketing instinct,' which the lab considers a virtue.

Tomoko works primarily from Kyoto and participates in the world-models cross-cell review pool through asynchronous artefact review rather than synchronous calls. Her review notes are widely read across the axis.

Background

Ph.D. mathematics, Kyoto University, 2014. Visiting scholar at MIT (Math), 2018-2019.

Prior to alphabell: Kyoto University; Riken AIP collaborator (2018-2021); Volterra Cognition.

Selected publications

Full publications index →

Recent talks

  • Operators learn what they are told to learn, NeurIPS 2024
  • Cross-modal coherence at scale, ICML 2025
Working with

Tomoko is currently part of node-cell voronoi-19, working under the World models research axis. The cell is open to substantive correspondence from researchers working on adjacent problems; route requests through voronoi-19@alphabell.com or directly to Tomoko at tomoko-niwa@alphabell.com.