Research map / Large language models / Agents, code and math
TritonGym: A Benchmark for Agentic LLM Workflows in Triton GPU Code Generation
Yue Guan, Yichen Lin, Xu Zhao, Jianzhu Yao, Xinwei Qiang, Zhongkai Yu, Pramod Viswanath, Yufei Ding, Adnan Aziz
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 |
|---|---|---|---|
| use AlphaEvolve workflow instead of one-shot generation | Claude-4.5 · TritonGym full benchmark | Pass@1: 31.6% -> 54.8% | Improved |
| use AlphaEvolve workflow instead of one-shot generation (average over LLMs) | LLM backbones · TritonGym full benchmark | Pass@1: 32.8% -> 42.3% | Improved |
| use Geak workflow instead of one-shot generation (average over LLMs) | LLM backbones · TritonGym full benchmark | Pass@1: 32.8% -> 36.7% | Improved |
| use Geak workflow instead of one-shot generation | Claude-4.5 · TritonGym full benchmark | Pass@1: 31.6% -> 33.9% | Mixed |
| use AlphaEvolve workflow instead of one-shot generation | Claude-4.5 · TritonGym OOD samples | Pass@1: 23.1% -> 61.5% | Improved |
| use AlphaEvolve workflow instead of one-shot generation | Claude-4.5 · TritonGym full benchmark | Perf@1: 0.307 -> 0.583 | Improved |
| use Geak workflow instead of one-shot generation | GPT-5 · TritonGym full benchmark | Pass@1: 23.9% -> 28.7% | Improved |
| use Geak workflow instead of one-shot generation | Claude-4.5 · TritonGym full benchmark | Perf@1: 0.307 -> 0.321 | Mixed |
| use Geak workflow instead of one-shot generation | DeepSeek-V3 · TritonGym full benchmark | Pass@1: 36.1% -> 38.1% | Improved |
| use AlphaEvolve workflow instead of one-shot generation | Llama-4 · TritonGym full benchmark | Pass@1: 24.6% -> 35.1% | Improved |
| use AlphaEvolve workflow instead of one-shot generation | Qwen-3 · TritonGym full benchmark | Perf@1: 0.282 -> 0.358 | Improved |
| use AlphaEvolve workflow instead of one-shot generation | LLM backbones (average) · TritonGym DSL samples | Pass@1: 14.6% -> 25.0% | 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 0.89)
- EGG: An Expert-Guided Agent Framework for Kernel Generation (ICML 2026 · similarity 0.88)
- The Automated LLM Speedrunning Benchmark: Reproducing NanoGPT Improvements (NeurIPS 2025 · similarity 0.88)
- Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.87)
- CUDABench: Benchmarking LLMs for Text-to-CUDA Generation (NeurIPS 2026 · similarity 0.87)
- EvoCodeBench: Evaluating Coding Agents in Multi-Turn Iterative Interactions (NeurIPS 2026 · 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.
