Research map / Large language models / Agents, code and math
QiMeng-Kernel: Macro-Thinking Micro-Coding Paradigm for LLM-Based High-Performance GPU Kernel Generation
Xinguo Zhu, Shaohui Peng, Jiaming Guo, Yunji Chen, Qi Guo, Yuanbo Wen, Hang Qin, Ruizhi Chen, Qirui Zhou, Ke Gao, Yanjun Wu, Chen Zhao and 1 more
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 |
|---|---|---|---|
| add MTMC macro-thinking/micro-coding | Gemini 2.5 Pro · KernelBench Level 1 (H100) | execute accuracy (%): 63 -> 100 | Improved |
| add MTMC | Gemini 2.5 Pro · KernelBench Level 3 (H100) | mean speedup: 0.27 -> 0.77 | Improved |
| add MTMC | Gemini 2.5 Pro · KernelBench Level 3 (H100) | execute accuracy (%): 36 -> 70 | Improved |
| add MTMC | Gemini 2.5 Flash · TritonBench-T (A100) | execute accuracy (%): 9.04 -> 54.82 | Improved |
| add MTMC | Gemini 2.5 Flash · TritonBench-G (A100) | fast1 (%): 4.35 -> 9.78 | Improved |
| add MTMC | Gemini 2.5 Flash · TritonBench-G (A100) | fast2 (%): 1.63 -> 1.63 | No clear effect |
| generate stepwise with Micro Coding instead of single pass (ablation) | Gemini 2.5 Flash · KernelBench Level 1 | execute accuracy (%): 60% -> 94% | Improved |
| generate stepwise instead of single pass (ablation) | Gemini 2.5 Flash · KernelBench Level 3 | execute accuracy (%): 10% -> 64% | Improved |
| use RL-trained policy (DeepSeek-Coder-1.3B) instead of GPT-4o for macro thinking (ablation) | MTMC · KernelBench Level 3 | execute accuracy (%): 40% -> 100% | Improved |
| use hardware-aware action space for GPT-4o macro thinking (ablation) | GPT-4o without policy · KernelBench Level 3 | execute accuracy (%): 20% -> 40% | Improved |
| use Llama-3.2-1B instead of DeepSeek-Coder-1.3B as policy (ablation) | MTMC · KernelBench Level 3 | execute accuracy (%): 100% -> 80% | Hurt |
| generate CUDA instead of Triton for matmul kernels (ablation) | MTMC · KernelBench matmul tasks | execution time (ms): task 6: 4.43 -> 1.34; task 2: 1.36 -> 1.66 | Mixed |
Closest papers by meaning
- 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float (DFloat11) (NeurIPS 2025 · similarity 0.88)
- EGG: An Expert-Guided Agent Framework for Kernel Generation (ICML 2026 · similarity 0.88)
- Practical Offloading for Fine-Tuning LLM on Commodity GPU via Learned Sparse Projectors (AAAI 2025 · similarity 0.88)
- KernelBench: Can LLMs Write Efficient GPU Kernels? (ICML 2025 · similarity 0.88)
- CUDABench: Benchmarking LLMs for Text-to-CUDA Generation (NeurIPS 2026 · similarity 0.88)
- EasySpec: Layer-Parallel Speculative Decoding for Efficient Multi-GPU Utilization (NeurIPS 2025 · similarity 0.87)
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.
