Maxwell Adam
max at v3rv dot comPublications
Spectroscopy at Scale: Finding Interpretable Structure in Pythia-1.4B
Susceptibilities as an interpretability technique at 1B+ scale: show and tell from Pythia-1.4B.
Daniel Murfet*, Andrew Gordon*, Maxwell Adam*, George Wang*, Jesse Hoogland*, Garrett Baker*, William Snell, Stan van Wingerden, Adam Newgas, Billy Snikkers, Rohan Hitchcock*
How to Scale Susceptibilities
The mathematics and methodologies behind scaling and analyzing susceptibilities.
Andrew Gordon*, Maxwell Adam*, Jesse Hoogland*, Daniel Murfet*
The Loss Kernel: A Geometric Probe for Deep Learning Interpretability
We introduce the loss kernel, an interpretability method for measuring similarity between data points according to a trained neural network. The kernel is the covariance matrix of per-sample losses computed under a distribution of low-loss-preserving parameter perturbations.
Maxwell Adam*, Zach Furman*, Jesse Hoogland
Bayesian Influence Functions for Hessian-Free Data AttributionICLR 2026
We propose the local Bayesian influence function (BIF), an extension of classical influence functions that replaces Hessian inversion with loss landscape statistics that can be estimated via stochastic-gradient MCMC sampling.
Philipp Alexander Kreer*, Wilson Wu, Maxwell Adam, Zach Furman, Jesse Hoogland
Influence Dynamics and Stagewise Data AttributionICLR 2026
We introduce a framework for stagewise data attribution grounded in singular learning theory. We predict that influence can change non-monotonically, including sign flips and sharp peaks at developmental transitions. We first validate these predictions in a toy model, then at scale in language models, where token-level influence changes align with known developmental stages.
Jin Hwa Lee*, Matthew Smith*, Maxwell Adam, Jesse Hoogland
Interpreting Complexity
Leveraging differences in circuit complexity to uncover and interpret latent structure in neural networks, using measures from Singular Learning Theory.
Maxwell Adam
Experience
Timaeus — Research Scientist
July 2025 – present
Timaeus — Research Intern
Jun 2025 – July 2025
Zendesk, Melbourne Aus — Engineering Intern
September 2022 – March 2023
Education
The University of Melbourne
February 2024 – Present
Mix of neuroscience, math, and computer science
RMIT University, Melbourne Aus
February 2023 – November 2023
Bachelor of CS