Yuki Cho

Research

Yuki works on the disagreement protocol — what happens when a paired interpretability cell and a producing cell read the same training logs and reach different conclusions about whether a run should continue. Her axis-internal note on disagreement handling has become required reading for newly-paired cells.

She has been with alphabell since 2018, joining from a Stanford postdoc in the HAI group. She is one of the two interpretability-cell leads who called the halt on the 25-19 RSI run; the delayed-release report for that halt is co-authored with Liora Sabatini.

Yuki splits her time between Tokyo and the Hong Kong anchor; the East Asia time-zone overlap with the lab's other Hong Kong and Singapore contributors is part of what makes the paired-cell rhythm work for her cell.

Background

Ph.D. computer science, University of Tokyo, 2015. Postdoc at Stanford (HAI), 2015-2017.

Prior to alphabell: University of Tokyo; Constellation; Helios Safety Group.

Selected publications

• Sep 2024 · ab-interpretabili
  Interpretability Cell Pairing: how every dual-use capability run gets a watchful sibling
  Karima Belkadi, Hester Vandekerckhove, Yuki Cho
• Jun 2025 · ab-recursive-modi
  Modification-Under-Review: protocols for safe self-modification of training procedures
  Liora Sabatini, Yuki Cho, Aravind Periyasamy
• May 2025 · ab-mechanistic-ci
  Mechanistic Circuit Analysis at Frontier Scale: cells as a unit of interpretability
  Jiang Yifei, Nico Almgren, Karima Belkadi, Hester Vandekerckhove
• Sep 2025 · ab-scalable-overs
  Scalable Oversight for Multi-Step Agent Systems: a Debate-Plus-Trace Approach
  Ifeoma Nwosu-Howard, Hiroshi Tanigawa, Maral Lotfi, Ruth Wernicke
• Nov 2024 · ab-sandboxed-self
  Sandboxed Self-Modification: a confinement specification and implementation
  Liora Sabatini, Cheung Wai-Lin, Marek Holub

Full publications index →

Recent talks

• How an interpretability cell calls a halt, AISI Pairing Workshop 2025
• Reading traces under disagreement, NeurIPS workshop 2024