research
The work is its own reward. — Sir Arthur Conan Doyle, British writer and physician
This page collects the research I have been lucky enough to take part in, from my first independent paper to the labs I am working with today. Most of it orbits the same question: how can we build machines that learn the way people do, incrementally, hierarchically, and with structure we can actually read? Each card shows when the work ran (ongoing efforts come first, and are marked as running to the present), the lab it lived in, what came out of it, and links to the code and papers behind it - please reach out if any of it sounds interesting!
Persona Conditioning
with Dr. Levin Brinkmann
Expert-token conditioning is a training protocol I have developed with Dr. Levin Brinkmann that specializes a single shared backbone on a bank of per-expert tokens instead of per-expert FFNs, alternating between fitting the tokens against a frozen backbone and training the backbone with the tokens held fixed - it is capacity-free and addressable, making it a lightweight alternative to Mixture-of-Experts routing.
We are currently working to extend the protocol from synthetic and semantic domains to real speakers, treating each token as a retrieval key for a persona learned from diarized speech transcripts. Our final goal is not only to prove the efficiency of the technique, but also to justify methods of developing human-like creativity in LLMs, which tend to approximate the mean but not the standard deviation.
outputs/
- scaling the method to a more expansive dataset of real speakers
- in preparation for ICLR 2027
TRELLIS
with Dr. Pat Langley
Cobweb is a symbolic algorithm that aggregates instances into “concepts” and forms a hierarchy over these instances - it is unsupervised, piecemeal, and incremental, making it a comprehensive model of human learning. I am currently working with Dr. Pat Langley to implement the Cobweb algorithm in a proposed theoretical framework to model psychological “chunking” for efficient parsing, storage and retrieval of concepts based on a set of defined relations.
In completing this work, I also documented psychological findings prevalent in Cobweb: we presented Cobweb as a theory of categorization refined on an information-theoretic rational analysis, and investigated how well it can account for key findings from classic categorization experiments conducted by Hayes-Roth and Hayes-Roth (1977), Medin and Schaffer (1978), and Smith and Minda (1998).
outputs/
- presented at the Computation and Cognition Conference (2026)
- a paper on Cobweb as a theory of categorization, accepted to CogSci 2026
- working towards a finalized theory for Advances in Cognitive Systems 2026
- an interactive explainer of the algorithm on the education page
Cobweb in Retrieval and Summarization of Language
with Dr. Chris Maclellan
(see notes on Cobweb in my work with ISLE) I am currently leading an independent study under Dr. Chris Maclellan in creating a robust, industry-sustainable neurosymbolic solution to utilizing language in agentic systems, using Cobweb as the primary algorithm. Our work takes on two main applications to modern semantic embedding analysis.
Our first approach is retrieval-aligned - we show that Cobweb performs as a strong search algorithm for dense vector representations, matching performance of traditional systems while offering strong intermediary prototypes. Our more recent approach aimed to leverage the clustering nature of Cobweb in a field that benefits from both the incremental and hierarchical nature of Cobweb - topic-modeling.
outputs/
- a preliminary study at Advances in Cognitive Systems 2025, qualifying Cobweb as a search algorithm and showing its ability to utilize different data forms
- Cobweb as a topic model sets a new SOTA in the hierarchical and incremental fields, a sustainable way to continuously track topic evolutions (accepted to ACL Findings 2026)
Encoding Models & Topographic Constraints
with Prof. Anya Ivanova
I joined Professor Anya Ivanova and began contributing to the development of encoding models, models that can convert latent content that represents stimuli into direct voxel representations of brain states, measured by fMRI data. Over the course of the project, I worked on creating a novel architecture that combined predictions from various layers of an LLM via attention to best predict brain activations.
I also developed a research project furthering topographic constraints in neural architectures for gains in pruning performance, and to pass on some of the blessings of continual learning to all architectures - specifically, how topography can improve diversity and sparsity constraints to reduce catastrophic forgetting and improve diversity of activations.
outputs/
- preprocessing decisions in the pipeline had strong effects as to which layer provided the best correlation
- syntactic decisions were localized in earlier layers and the auditory cortex, whereas semantics occurred in the middle to later layers and the language network
SimplyPhi
with Dr. Larissa Albantakis
I did volunteer work for Dr. Larissa Albantakis on a theory called Integrated Information Theory, an information-theoretic framework that attempted to calculate the distributed “consciousness” of a system. I used IIT to evaluate the performance of neural networks as a saliency test, identifying the most relevant parts of the network under the “substrate of consciousness” defined by the framework.
While IIT is an extremely assuming framework to attribute to consciousness, as a metric of “usefulness” by analyzing the distributed information across a system, it has a lot of play, and I look forward to seeing information theory take greater prominence in shaping the field of ML interpretability.
AI-for-Advertising
outputs/
- paper accepted to the Applied Human Factors and Ergonomics Conference (unfortunately, I was unable to attend due to another commitment)
BiasWrappers
outputs/
- a paper on TechRxiv
- a Python package that currently stands at ~32k downloads