Hiroshi Tanigawa

Research

Hiroshi co-authored the debate-plus-trace oversight result that adapts debate-based oversight to multi-step agent execution. The result took close to two years of work between his cell (lebesgue-22) and the agentic axis's substrate team; it is a canonical example of cross-axis collaboration at alphabell.

He is now leading the v2 of the debate-plus-trace protocol, which targets agents operating in environments where the evaluator is up to 10× weaker than the action-taker (vs. 6× in the published v1 result). The v2 work is in active development; cell-internal progress reports suggest the gap can be narrowed further.

Hiroshi is based in Osaka. He maintains a long-running set of public lecture notes on scalable oversight that several Japanese-speaking university courses now use as a reference.

Background

Ph.D. computer science, University of Osaka, 2015. Postdoc at CMU (LTI), 2015-2017.

Prior to alphabell: CMU LTI; Constellation; Praxis AI Studies.

Selected publications

• 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
• Sep 2024 · ab-interpretabili
  Interpretability Cell Pairing: how every dual-use capability run gets a watchful sibling
  Karima Belkadi, Hester Vandekerckhove, Yuki Cho
• Mar 2025 · ab-verifiable-pol
  Toward Formal Verification of Learned Policies in Bounded Environments
  Aviva Stern, Sun Kyung-min, Felipe Avelar
• 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
• 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

• Debate-plus-trace in production, NeurIPS 2025 (oral)
• Oversight as an engineering problem, ICML 2024