Rohan Hitchcock

I am a research scientist at Resolution, a nonprofit AI alignment research institute, where I mostly work on singular learning theory. I completed my PhD at the University of Melbourne in 2026.

Papers

• “Singular Learning Theory for Deep Learning Interpretability: Stochastic Sampling Methods and Applications to Learning from Physical Systems.” PhD thesis, University of Melbourne, 2026. thesis
• “Susceptibilities for Neural Networks Learning from Physical Data.” R. Hitchcock, G. W. Delaney, J. H. Manton, R. Scalzo, J. Zhu. NeurIPS 2026. arXiv
• “From Global to Local: A Scalable Benchmark for Local Posterior Sampling.” R. Hitchcock, J. Hoogland. 2025. arXiv
• “Emergence of Computational Structure in a Neural Network Physics Simulator.” R. Hitchcock, G. W. Delaney, J. H. Manton, R. Scalzo, J. Zhu. XAI4Science workshop, ICLR 2025. arXiv

Research notes

• “Interpretability as a Benchmark for Sampling.” A. Gordon, R. Hitchcock, I. Mouratidis, J. Wind, D. Murfet. Resolution, September 2026. report
• “Guide for Sampling Hyperparameter Selection.” R. Hitchcock, J. Hoogland, G. Wang, A. Gordon. Timaeus, April 2026. report
• “Interpreting the Ising Model.” A. Gordon, R. Hitchcock, D. Murfet. Timaeus, April 2026. report
• “Spectroscopy at Scale: Finding Interpretable Structure in Pythia-1.4B.” D. Murfet, A. Gordon, M. Adam, G. Wang, J. Hoogland, G. Baker, W. Snell, S. van Wingerden, A. Newgas, B. Snikkers, R. Hitchcock. Timaeus, April 2026. report

Talks

• “Sampling Hyperparameter Selection in SLT.” ILIAD, Berkeley, August 2026.
• “Towards Higher Confidence in Alignment.” Australian AI Safety Forum, Sydney, July 2026. video
• “The Empirical Process in Singular Learning Theory.” SLT Seminar, online, September 2025. slides · video
• “Local Learning Coefficient Estimation: The Story So Far and Open Problems.” ILIAD, Berkeley, August 2025. slides
• “Induction Heads and Phase Transitions.” SLT & Alignment Summit, Berkeley, June 2023. slides · video

Earlier work

• “Differentiation, Division and the Bicategory of Landau-Ginzburg Models.” MSc thesis, University of Melbourne, 2022. thesis · code · talk
• “Display of Native Antigen on cDC1 That Have Spatial Access to Both T and B Cells Underlies Efficient Humoral Vaccination.” Kato et al. Journal of Immunology, 2020. paper