Personal research journal
Wesley Lu
Welcome to My Page
I’m an undergraduate . My interests lie at the intersection of machine learning, language models, and performant systems, with a long-term focus on LM/ML research. I currently work with at the University of Virginia, developing machine learning systems for decision-making. I also conduct research on post-training techniques and ai benchmarking in the at the University of Illinois Urbana-Champaign.
I spent two summers at Amazon developing backend infrastructure for a conversational AI product. My work involved exploring chain-of-thought reasoning, agentic interactions, and LangGraph architectures to improve system reliability and performance. Before Amazon, I developed websites for the real estate industry.
Outside of computer science, I have experience in fundamental equity research, personal portfolio management, and reselling ventures, through which I generated more than six figures in revenue.
Viewed by 1,319 humans
Recent
Every Interface Is Now an Agent Interface: Computer Use Agents
A short writeup on GUI testing for computer use agents and motivations for creating Inspector
Training Verifiers Through Debate: Three-Agent Prover-Verifier Game
A three-agent extension of the prover-verifier game where two adversarial provers debate to train a more robust verifier.
ML Undergraduate Researcher · University of Virginia
Under the guidance of Hongru Du, PhD, I’m building an LM for human decision modeling pipeline that generates interpretable decision rules from primitives and iteratively improves them using model feedback. I also incorporate prior probabilities and update them with observed outcomes on a mobility dataset, using gradient descent to better fit parameters and calibrate predicted choice probabilities.
Jump's Crypto Reading List (ip)
This is a collection of notes from Jump's Crypto Reading List
Software Development Engineer Intern · Amazon
Incoming 2026 summer
Pattern-Driven Evaluation of LLM Code Optimization: Isolating Compiler-Transcendent Semantic Reasoning
Existing evaluations of LLM-based code optimization primarily measure end-to-end speedup, obscuring whether a model actually performs the intended optimization. For compiled languages, this is especially problematic: observed gains may reflect transformations already recoverable by modern optimizing compilers rather than optimization behavior attributable to the model itself. We present PatternBench, a benchmark for studying compiler-transcendent optimization, i.e., performance improvements that remain after aggressive compiler optimization. PatternBench targets residual optimization opportunities across seven classes of transformations and pairs runtime evaluation with an AST-based optimization faithfulness analysis that verifies whether the intended optimization is realized in the generated program. Experiments with frontier, open-weight, and fine-tuned models show that runtime speedup alone provides an incomplete account of optimization capability. Optimization faithfulness is strongly associated with realized speedup, and fine-tuning improves both faithfulness and optimization effectiveness. These results suggest that evaluation of LLM-based code optimization should measure not only performance outcomes, but also whether the target transformation was successfully executed.