Zach does mechanistic interpretability

• I’m working on independent mech interp research these days! Feel free to reach out about collaboration.
• My interests are broadly in (1) sparse representations + structured dependence, (2) interp methods useful for biological models, and (3) scalable interp approaches.

Contact Me

hey@zachmaas.com ⋅ @zmaas

Highlighted Work:

• Conditional Dependence Structure in Sparse Autoencoder Features: ICML Mech Interp Workshop 2026 spotlight. Dependence graph structure in SAEs
• Graph Structure Compresses Contrastive Neuron Steering Targets: Can we steer LLM behavior using only MLPs + structured dependence?
• Preliminary Results on Graph SAEs: Building dependence graphs off of SAEs to understand their structure
• Research Notes: Sequence to Read Models: Notes on how to build (and not build) sequence-to-read genomic models

Papers I Liked Recently:

• Subliminal Learning Is Steering Vector Distillation
• Language Model Circuits Are Sparse in the Neuron Basis
• Do Sparse Autoencoders Capture Concept Manifolds?
• Targeted Neuron Modulation via Contrastive Pair Search