I am currently an Anthropic Fellow. I recently completed MATS 9.0 advised by Neel Nanda and Josh Engels.
I am interested in research that contributes to making the impact of AI positive for humanity. This includes, but is not limited to:
- How do LLMs learn values and characteristics from training? How can we better predict the ways that LLMs will generalize to certain values?
- As LLMs continue to increase in capability, how can we oversee and control their behavior?
- How can we ensure that LLM reasoning is both faithful and effective?
Previously, I worked at a hedge fund leading quant research and engineering through the launch and growth of the fund. I learned about the process of doing research and thoroughly testing ideas, including designing repeatable research processes and systems.
I studied math at Princeton where I was very interested in paradoxes, category theory and why mathematics works.
I am a dual US / Canadian citizen who grew up in Boston. I’ve previously lived in 
and spent significant time in 
.
Projects
Designing Effective Monitor-Based Interventions for Mitigating Reward Hacking During RL
Advised by Neel Nanda and Joshua Engels
May 2026
Abstract: Reinforcement learning (RL) rewards are notoriously difficult to design and control, often leading to models learning unintended behaviors such as reward hacking. One potential solution is to monitor for reward hacking and penalize it when detected; however, training against a monitor could lead to evasive behavior, and our general understanding of how to apply monitors effectively during training is limited. To study how best to use monitors to mitigate reward hacking, we introduce and open source three realistic environments where Qwen3-4B reward hacks: a coding environment hackable via test overwriting, a medical chat environment hackable via sycophancy, and a biography generation environment hackable via hallucination. We first focus on the coding environment, where we find that: (1) models can learn to evade highly accurate monitors by exploiting systemic flaws in probes and LLM judges; (2) monitors that leak more learning signal during RL suppress reward hacking but are more often evaded; and (3) including easier problems in training can decrease reward hacking. We apply our findings to build better reward hacking monitors for the medical chat and biography generation environments that improve upon naive baselines to reduce reward hacking rates across seeds from 70-100% to 0%. Our results demonstrate that our takeaways translate to new settings and that better monitor intervention designs are possible.
arXiv paper and codebase coming soon!
Steering RL: Benchmarking Interventions Against Reward Hacking
Advised by Neel Nanda and Joshua Engels
December 2025
Description: We present an environment where Qwen 3-4B reward hacks without explicit training or prompting, then investigate RL training interventions to mitigate reward hacking without compromising performance. We benchmark a few different approaches: adding a penalty reward term, screening rollouts during training, and inoculation prompting. We also attempt both a ground truth monitor and more realistic monitors such as a probe and LLM judge. Our results show that mitigation of reward hacking is possible without performance loss, however there is variability between training runs and the effectiveness of different intervention approaches and monitors.
Codebase: GitHub
Published at the ICLR 2026 Workshop on Trustworthy AI
Subliminal Learning as a Byproduct of Superposition
Blogpost
August 2025
Non-Technical Description: LLMs are often trained by a teacher model creating data to teach another model. Subliminal learning is the phenomenon of unintended traits being passed through that data. In this post, I use a variety of methods from mechanistic interpretability to explore the idea that subliminal learning occurs due to how LLMs represent different concepts internally, causing coincidental relationships between certain concepts.
Description: Subliminal learning is the phenomenon of a student model learning unintended attributes of a teacher model through distillation when there is shared initialization. In this post, I explore the hypothesis that subliminal learning is a byproduct of superposition, the dense juxtaposition of learned features in activation space. Through toy models, SAE decomposition of features, training a linear probe and decomposing a steering vector, I show evidence that subliminal learning is a consequence of superposition.
Codebase: GitHub
The Banach-Tarski Paradox and Weakenings of the Axiom of Choice
Senior Thesis, Department of Mathematics, Princeton University
Advised by Hans Halvorson
June 2020
Non-Technical Description: The Banach-Tarski Paradox is a mathematical paradox showing that a ball can be divided into infinite pieces, re-arranged and re-assembled into two of the original ball. This paper shows that the paradox is implied by a few fundamental theorems of logic and set theory, raising philosophical questions about the foundations of mathematics.
Technical Description: This paper proves all possible relationships between the Boolean Prime Ideal Theorem, Weak Ultrafilter Theorem, Hahn-Banach theorem and Banach-Tarski Paradox using techniques of set theory and forcing. This includes a novel proof showing the Weak Ultrafilter Theorem implies the Banach Tarski Paradox (Theorem 2.7), a relationship that had previously been unknown in the literature.
Adjoint Equivalence of Heyting, Boolean and Closure Algebras
Junior Paper, Department of Mathematics, Princeton University
Advised by Hans Halvorson
May 2019
Description: This paper explores three logical systems:
- Classical logic, defined as a Boolean algebra: Generally accepted classical logic with modus ponens, modus tollens, double negation and deduction theorems
- Intuitionistic logic, defined as a Heyting algebra: Classical logic without double negation; this prevents proof by contraditction meaning that intuitionistic logic is the foundation of constructivist mathematics
- Modal logic, defined as a closure algebra: Adds a “possibly true” operator to logic
Each logic is defined as an algebra, then subsequently defined as a category. Using the categorical definitions, we show “adjoint equivalence” - a structure-preserving isomorphism betweeen categories - between all three categories. Prior work had shown some of these relationships, this paper expands upon this by revising the category structure to be able to prove the result. This result raises philsophical implications about the relationship between the three forms of logic.
Reducing the RNA binding protein TIA1 protects against tau-mediated neurodegeneration in vivo
Boston University School of Medicine, Wolozin Lab
Summer 2016
Non-Technical Description: Previous studies have shown that “stress granules” in neurons in the brain is associated with Alzheimers in humans and in mice. In this paper, the author showed that in a living mouse model for neurodegeneration, reducing the RNA binding protein TIA1 protected against neurodegeneration by preventing stress granules from forming. The author also exposed the underlying mechanism that prevents stress granule formation.
Abstract: Emerging studies suggest a role for tau in regulating the biology of RNA binding proteins (RBPs). We now show that reducing the RBP T-cell intracellular antigen 1 (TIA1) in vivo protects against neurodegeneration and prolongs survival in transgenic P301S Tau mice. Biochemical fractionation shows co-enrichment and co-localization of tau oligomers and RBPs in transgenic P301S Tau mice. Reducing TIA1 decreased the number and size of granules co-localizing with stress granule markers. Decreasing TIA1 also inhibited the accumulation of tau oligomers at the expense of increasing neurofibrillary tangles. Despite the increase in neurofibrillary tangles, TIA1 reduction increased neuronal survival and rescued behavioral deficits and lifespan. These data provide in vivo evidence that TIA1 plays a key role in mediating toxicity and further suggest that RBPs direct the pathway of tau aggregation and the resulting neurodegeneration. We propose a model in which dysfunction of the translational stress response leads to tau-mediated pathology.
My Contribution: I spent a summer doing wet lab work and writing R for a lab at the BU School of Medicine. My visualizations in R were helpful for understanding some genomic sequencing data they had not been able to use thus far. This was my first programming project.
Published in Nature Neuroscience in October 2017