Jikai Jason Li

Honours Bachelor of Science in Computer Science · Toronto, Ontario

• Dean’s List Scholar.
• Coursework: Machine Learning (CSC311), Databases (CSC343), and Software Engineering (CSC301).

Department of Computer Science · Prof. Fan Long

LLM Compiler Superoptimization

• Built a benchmark pipeline for an LLM code optimizer and extended experiments to multi-file C/C++ programs, validating 139 cases across 49 LLVM benchmark families, including the 87-file SPASS theorem prover.
• Ran 400 generated optimization attempts across 11 workloads using 16 parallel search workers and four verifier agents; automated correctness verification, failure handling, and result collection in AWS S3.
• Reduced later verification prompt size by approximately 75% in a SPASS experiment by sending new evidence incrementally.
• Used verifier counterexamples to revise optimization constraints, improving targeted aggregate speedup from 1.07× to 1.24× while preserving correctness.

LLM-Driven GPU Kernel Optimization (CuTeGen)

• Third author on a NeurIPS 2026 submission on LLM-driven GPU kernel optimization using CUDA/CuTe; evaluated generated kernels across 209 KernelBench tasks on remote NVIDIA GPUs.
• Built tooling for correctness checks, CUDA performance timing, token tracking, and recovery of failed optimization runs.
• Debugged CUDA/CuTe kernels and PyTorch extensions, tracing compilation failures, GPU crashes, and incorrect outputs to precision, tensor-shape, and implementation errors.

Shipped TypeScript and React features for the CSSU website, serving a student union representing more than 5,000 students, and collaborated in a five-person team through GitHub pull requests and code reviews.

• Built a decoder-only Transformer in PyTorch with causal multi-head attention, positional embeddings, autoregressive generation, supervised fine-tuning, checkpointing, and device-aware training.
• Developed an RLHF pipeline with a learned reward model, frozen reference policy, value critic, batched rollouts, GAE, token-level KL regularization, and clipped PPO updates.
• Implemented a 500-token BPE tokenizer that compressed a 1.1-million-character corpus to 574,000 tokens (1.94×), with tools for analyzing policy likelihoods, reward-model failures, reward hacking, and exploration.