Taku Ito

Bio

I'm a research scientist in the mathematics and optimization group, working primarily in machine learning, compositionality, and quantum error correction.

Link to personal webpage: ito-takuya.github.com Link to scholar: https://scholar.google.com/citations?user=Ym08_RMAAAAJ&hl=en

Publications

Transformer Circuits Can Realize Clustering Algorithms
- - Kenneth Clarkson
  - Lior Horesh
  - et al.
- 2026
- ICML 2026
Conference paper
Task Representational Dynamics for Compositional Generalization During Context-Dependent Cognitive Tasks
- - Lakshman Chakravarthula
  - Michael Cole
  - et al.
- 2026
- ICLR 2026
Workshop paper
Specifying exact circuit algorithms in universal transformers
- - Taku Ito
  - Ruchir Puri
  - et al.
- 2025
- NeurIPS 2025
Workshop paper
Learning interpretable positional encodings in transformers depends on initialization
- - Taku Ito
  - Luca Cocchi
  - et al.
- 2025
- ICML 2025
Workshop paper
Geometry of naturalistic object representations in recurrent neural network models of working memory
- - Xiaoxuan Lei
  - Taku Ito
  - et al.
- 2024
- NeurIPS 2024
Conference paper
On the generalization capacity of neural networks during generic multimodal reasoning
- - Taku Ito
  - Soham Dan
  - et al.
- 2024
- ICLR 2024
Conference paper

Blog posts

Understanding AI through the algorithms they compute
Technical note
Takuya Ito
18 Aug 2025
- AI

Top collaborators

Parikshit Ram

Research Staff Member

Murray Campbell

Distinguished Research Scientist, Mathematics of Computation

James Kozloski

Principal Research Scientist - Biomedical AI

Tim Klinger

Senior Research Scientist, Math and Algorithms