Research map / Reinforcement learning / RL from feedback
Dr. Kernel: Reinforcement Learning Done Right for Triton Kernel Generations
Wei Liu, Jiawei Xu, Yingru Li, Longtao Zheng, Tianjian Li, Qian Liu, Junxian He
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 GRPO mean baseline with TRLOO leave-one-out baseline (ablation) | multi-turn RL for Triton kernels · KernelBench Level 2, Qwen3-8B-Base | Fast@1 | Improved |
| train multi-turn instead of single-turn (ablation) | RL with TRLOO · KernelBench Level 2, Qwen3-8B-Base | Fast@1 | Improved |
| use reward-to-go return (gamma=1) instead of gamma=0 (ablation) | TRLOO multi-turn RL · KernelBench Level 2 | Fast@1 first turn | Improved |
| add execution-based hacking check to the environment (ablation) | KernelGYM RL training · KernelBench Level 2 validation | Fast@1 under hacking check | Improved |
| add mismatch rejection sampling (MRS) (ablation) | TRLOO multi-turn RL · KernelBench Level 2 | training stability and Fast@1.2 | Mixed |
| add profiling-based rewards (PR) and profiling-based rejection sampling (PRS) on top of MRS (ablation) | TRLOO multi-turn RL · KernelBench Level 2 | Fast@1.2 | Improved |
| multi-turn RL after cold-start SFT | Qwen3-8B-Base · KernelBench Level 2 | Fast@1.2: 5.6 -> 20.0 | Improved |
| add sequential test-time scaling with context management | Dr. Kernel-14B · KernelBench Level 2 | Fast@1.2: 25.6 -> 31.6 | Improved |
| select best turn across history instead of last turn | Dr. Kernel-14B-STTS · KernelBench Level 2 | Fast@1.2: 31.6 -> 47.8 | Improved |
| context management (top-4 turns by reward) instead of vanilla extrapolation (ablation) | test-time scaling · KernelBench Level 2 | Fast@1.2 over turns | Mixed |
| evaluate under torch.compile baseline | Dr. Kernel-14B · KernelBench Level 2 | Fast@1.2: 25.6 -> 1.9 | Hurt |
| scale to Level 3 with the same training | Dr. Kernel-14B · KernelBench Level 3 | Fast@1.2: 1.2 vs GPT-5 12.0 | No clear effect |
Closest papers by meaning
- DRTriton: Large-Scale Synthetic Data Driven Reinforcement Learning for Triton Kernel Generation (NeurIPS 2026 · similarity 0.90)
- Reward-Free Kernel-Based Reinforcement Learning (ICML 2024 · similarity 0.89)
- Kernel-Based Function Approximation for Average Reward Reinforcement Learning: An Optimist No-Regret Algorithm (NeurIPS 2024 · similarity 0.87)
- Do Not Let Low-Probability Tokens Over-Dominate in RL for LLMs (ICLR 2026 · similarity 0.85)
- Reinforcement Learning for Code Optimization (NeurIPS 2026 · similarity 0.85)
- Accelerating RL for LLM Reasoning with Optimal Advantage Regression (NeurIPS 2025 · similarity 0.85)
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.
