Ph.D. Candidate, UW–Madison Computer Sciences | Founder & Maintainer, OpenSAE
Website · CV · LinkedIn · Hugging Face · Email
I study how learned representations shape model behavior across interventions, combinations, and changing contexts. My work brings together theory, controlled interventions, open-source research infrastructure, and large-scale evaluation in mechanistic interpretability, LLM reasoning, AI safety, and AI for scientific discovery.
My selected publications include Spectral-Spatial Interpretation (NeurIPS 2026), SegJointGene (Bioinformatics 2026), and DeepInfo (ICML 2022 Spotlight).
At OpenSAE, I build open research infrastructure for understanding and evaluating language models, including sparse autoencoder analysis, controlled interventions, and reproducible benchmarks.
Recent recognition: NeurIPS 2026 Top Reviewer · ICML 2026 Silver Reviewer
|
🧱 SAEBase Research infrastructure for evaluating, understanding, and intervening on sparse autoencoder features in language models. Code · Research Lab · Dataset |
🧩 SAR Measures when interventions on SAE features combine additively and when feature interactions change model behavior. |
|
🔀 RSS Studies how prompt changes redistribute sparse feature activity, testing persistence across prompt variants and implications for safety-probe failures. |
🔥 Spectral-Spatial Interpretation Identifies which spectral channels matter at each spatial location for a model's prediction. Code · Paper · Playground |
Other projects · 🧭 SAE Input Localization · 🎯 SymSelect · ⏳ The Limits of Reward Guidance · 🧬 SpatialPheno · 🔬 DeepInfo
Research directions
- Faithful representations and mechanistic interpretability. Test when learned features support faithful explanations, localization, and interventions.
- Interpretable AI for scientific discovery. Connect model predictions to spatial, spectral, and molecular structures.
- Compositionality, intervention, and systematic reasoning in LLMs. Study when components and transformations can be combined, reused, or generalized.
- Reliable LLMs and AI safety. Analyze uncertain preferences, unsupported user-identity assumptions, and privacy risks.