Publications

The frontier of AI is set by how well software understands hardware.

We research agents that work the way expert performance engineers do, from profiler evidence to verified code on real silicon. KAI is built on this work.

Papers05
  1. arXiv preprint Sparse compilation

    SparseDitto: An Agentic Sparse Compilation Framework through Architecture-Aware Synthesis on GPUs

    Shiyang Li, Guangyan Sun, Jinwei Tang, Yanzhi Wang, Mingyi Hong, Caiwen Ding

    Sparse matrix computation performance on GPU depends on how representation and execution schedule match the input structure and target hardware. No single implementation consistently dominates across sparsity patterns, operators, and hardwares…

    Read paper
  2. arXiv preprint Benchmarks

    CUDAHercules: Benchmarking Hardware-Aware Expert-level CUDA Optimization for LLMs

    Shiyang Li, Zijian Zhang, Guangyan Sun, Yuebo Luo, Winson Chen, Yanzhi Wang, Mingyi Hong, Caiwen Ding

    Large language models show promise for automated CUDA programming, however even the strongest coding models may still fall short of expert-level, architecture-aware optimization. We introduce CUDAHercules, a benchmark that evaluates generated CUDA…

    Read paper
  3. arXiv preprint Debugging

    CUDABeaver: Benchmarking LLM-Based Automated CUDA Debugging

    Shiyang Li, Haoyang Chen, Mattia Fazzini, Caiwen Ding

    Debugging CUDA programs has long been challenging because failures often arise from subtle interactions among hardware behavior, compiler decisions, memory hierarchy, and asynchronous execution. Current evaluations of LLM-based CUDA programming…

    Read paper
  4. ICML 2026 End-to-end programs

    StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement Learning

    Shiyang Li, Zijian Zhang, Winson Chen, Yuebo Luo, Mingyi Hong, Caiwen Ding

    Modern machine learning workloads increasingly rely on GPUs, yet achieving high end-to-end performance remains challenging due to dependencies on both GPU kernel efficiency and host-side settings. Although LLM-based methods show promise on automated…

    Read paper Code
  5. arXiv preprint Kernel optimization

    CudaForge: An Agent Framework with Hardware Feedback for CUDA Kernel Optimization

    Zijian Zhang, Rong Wang, Shiyang Li, Yuebo Luo, Mingyi Hong, Caiwen Ding

    Developing efficient CUDA kernels is increasingly critical for AI applications such as large-scale LLM training. However, manual kernel design is both costly and time-consuming, motivating automatic approaches that leverage LLMs for code generation…

    Read paper Code