Research

We decompose LLM reasoning token-efficiency into truncation robustness, conditional correctness, and workload-/trace-quality-normalized verbosity, to show that efficiency rankings can diverge from accuracy while revealing distinct sources of wasted tokens (verbosity, looping, or logic errors).

A synthetic benchmark grounded in Cognitive Load Theory (CLT) that generates natural-language logic puzzles with independently tunable parameters to diagnose LLM reasoning bottlenecks.

Building on the CogniLoad benchmark, this work introduces a novel efficiency metric for LLMs—tokens generated per solved puzzle—to evaluate computational cost alongside accuracy.

Investigates the application of machine learning methods to forecast stock movements. First study to successfully apply Machine Learning to quantitative financial statement data.

Authors listed in alphabetic order. Please refer to my thesis instead with all details.