Research map / Reinforcement learning / RL from feedback
CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning
Xiaoya Li, Xiaofei Sun, Albert Wang, Jiwei Li, Chris Shum
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace contrastive RL with vanilla GRPO after stage 1+2 (ablation) | CUDA-L1 · KernelBench all levels (A100) | mean speedup: 2.41 -> 3.12 | Improved |
| add contrastive RL stage after SFT and self-supervised stages (ablation) | CUDA-L1 · KernelBench all levels (A100) | mean speedup: 1.36 -> 3.12 | Improved |
| add self-supervised learning stage after SFT (ablation) | CUDA-L1 · KernelBench all levels (A100) | mean speedup: 1.14 -> 1.36 | Improved |
| use bucket or island exemplar sampling instead of random sampling (ablation) | contrastive RL · KernelBench all levels (A100) | mean speedup: 2.14 -> 3.12 | Improved |
| island sampling versus bucket sampling (ablation) | contrastive RL · KernelBench all levels (A100) | mean speedup: 3.21 vs 3.12 | No clear effect |
| update model parameters (RL) instead of fixed-parameter evolutionary prompting (ablation) | DeepSeek-R1 / CUDA-L1 · KernelBench all levels (A100) | mean speedup: 1.41 -> 3.12 | Improved |
| add evolutionary contrastive prompting to vanilla foundation models (ablation) | DeepSeek-R1 · KernelBench all levels (A100) | mean speedup: 0.88 -> 1.41 | Improved |
| apply A100-trained kernels to other GPUs | CUDA-L1 · KernelBench all levels | mean speedup: 2.38x to 3.85x across H20, 3090, L40, H100 | Improved |
| compare against torch.compile baselines instead of default on Level 3 | CUDA-L1 · KernelBench Level 3 (A100) | mean speedup: 2.96 -> 1.62 | Hurt |
| add CUDA Graphs to LSTM kernel optimization (ablation) | CUDA-L1 LSTM kernel · KernelBench Level 3 Task 35 | speedup: 1.0x -> 2.77x-3.42x | Improved |
| add mathematical short-circuit to Conv3d task (ablation) | CUDA-L1 kernel · KernelBench Level 2 Task 38 | speedup: 1.0x -> 28.6x+ | Improved |
Closest papers by meaning
- StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement Learning (ICML 2026 · similarity 0.91)
- Kevin: Multi-Turn RL for Generating CUDA Kernels (ICLR 2026 · similarity 0.90)
- CUDABench: Benchmarking LLMs for Text-to-CUDA Generation (NeurIPS 2026 · similarity 0.89)
- DOGE-Train: Discrete Optimization on GPU with End-to-End Training (AAAI 2024 · similarity 0.85)
- From Weight-Based to State-Based Fine-Tuning: Further Memory Reduction on LoRA with Parallel Control (ICML 2025 · similarity 0.84)
- Beyond the Trial-and-Error Loop: Hybrid Projection and Automated Tuning for Distributed Training (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
