Cognitive science · AI concentration · Class of 2028
Research
I use computational methods to study cognition in brains and machines, and to ask whether the systems we’re building share the same constraints.
What I work on
Since January 2025 I’ve been a research assistant in Steven Frankland’s Mind, Brain, & Computation Lab, where I study working memory in large language models. LLMs can hold their whole input in context, so storage is rarely the bottleneck. What we test is how well they can manipulate that information. We adapt classic paradigms from cognitive psychology, like digit span, to find where that manipulation breaks down and compare it with human capacity limits.
More generally, I’m interested in linking the internal mechanisms of large language models to their behavior, using mechanistic analysis to explain what we observe in behavioral experiments. The bigger questions I keep coming back to are continual learning, AGI, and consciousness.
Papers
-
2026
Backward Digit Span Benchmarks Working Memory in LLMs
Diak, C. J., Nguyen, N., Tibbetts, J.R., Webb, T. W., & Frankland, S. M.
· CogSci 2026
Because LLMs keep their entire input in the context window, storage-based working memory limits should barely apply to them. Work in cognitive science suggests, however, that capacity limits can also come from representational interference during manipulation. We introduce backward digit span (BDS) as a benchmark for inference-time sequence manipulation and evaluate 15 OpenAI models on sequences of 2 to 500 digits. Forward span is near-perfect, but backward span degrades with sequence length. Reasoning-optimized models reach longer backward spans with similar degradation profiles, and context-window size doesn’t explain the effect (GPT-5 models with 400K-token windows outperform GPT-4.1 models with 1.05M). Backward span also correlates significantly with performance on text-based Raven’s Progressive Matrices and ARC-AGI-1, mirroring the classic human relationship between working memory and fluid intelligence.
Relevant classes
- Winter 2025 Information Theory in Cognitive Science
- Winter 2025 What is Consciousness?First-year writing seminar
- Spring 2025 Computational Methods in Cognitive Science
- Fall 2025 Theories of Consciousness