Research Assistant — University of Alberta
I study what's latent in minds and language models — psychometrics from embeddings, real-time EEG, and LLM interpretability tools.
Pillar 01 · Steering & structure
Opening up language models — sparse-autoencoder feature steering, and evolving vectors through embedding space to predict human behaviour.
A Qwen3.5-2B with a sparse-autoencoder “Rat Brain” dial — drag it up and watch one steered SAE feature (#26631) take over the model’s personality. An open-source, rat-themed take on Golden Gate Claude.
Chris Westbury, Devon Yanitski, Anna Brooker
Genetic programming evolves 'average word' vectors in embedding space to predict lexical-decision and naming reaction times — operationalizing the cognitive notion of 'likely need'.
Pillar 02 · Meaning as co-occurrence
Can the geometry of an embedding space stand in for human survey responses? Semantic Factor Analysis recovers latent psychological structure from item text alone.
50 Big-Five items embedded by Qwen3-Embedding-8B (4096-d), projected to 3D with UMAP on cosine geometry. Items cluster by factor — meaning recovered from wording alone.
Pillar 03 · Behavioural probes
Do language models choose like utilitarians, and does reasoning shift their judgements the way it shifts ours? Measuring when LLMs replicate human behaviour.
Devon Yanitski
Does the human dual-process finding — that reasoning increases utilitarian moral judgment — extend to LLMs? Across 122 open-source models run on a 40-item moral-dilemma battery, switching on reasoning raised utilitarian responding and lowered altruism and deontology: the human dual-process signature, reproduced in a system with no affect, kin, or evolved morality.
Devon Yanitski
122 open-source LLMs took an updated Cognitive Reflection Test (Meyer et al., 2024), each item answered twice — once on intuition, once with reasoning. Holding arithmetic ability fixed, engaging reasoning alone moves the score, supporting the claim that the CRT is separable from ability — a cleaner demonstration than human data, where CRT scores confound ability, disposition, and motivation.
Pillar 04 · Neural signals
Real-time EEG tooling — Muse Visualizer turns the raw waveforms and oscillatory band activity from a consumer headband into highly customizable, artistic live representations of the brain.
Playing the Council of Elrond theme (LOTR) on piano while my live brain activity renders in Muse Visualizer.
Pillar 05 · AI-assisted research
I am building the scaffolding that lets AI agents carry out autonomous psychology research — starting with how they ground themselves in the literature.
A full paper runs ~30k tokens, so a 250k-token context fits only about 8 whole papers before it overflows. You can’t just paste the library into the prompt — that’s why a RAG system retrieves the few passages that actually matter. AP-RAG chunks academic papers into passages of 300–800 tokens, letting it synthesize information from hundreds of papers in a single context window.
Different questions need different evidence — pinning down a single paper isn’t the same as surveying a whole field — so AP-RAG offers several retrieval modes.
Combine entity- and relation-centric retrieval, merge the results, then recover the supporting chunks through the graph.
Also — built for fun
Get in touch
Always happy to talk LLMs, EEG, psychometrics, or a strange side project.