mlsanityData to 7 Oct 2026

Multi-turn RL with Structural and Performance Aware Rewards for CUDA Kernel Generation

arXiv 2607.20908 ↗First posted 23 Jul 2026cs.LG cs.AIcited by 0

On the map: LLM reasoning → Reinforcement learning for reasoning

Abstract (arXiv)

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful technique to enhance the reasoning capacity of LLMs for optimized code generation. However, existing RLVR approaches primarily rely on outcome-based signals such as correctness and speedup, overlooking performance-critical structural properties of programs that are essential for generating optimized code. In this work, we propose CudaPerf, a reflective RL framework that incorporates both verifiable execution rewards and structural code-aware rewards derived from parallelization features (e.g., memory coalescing, occupancy, Arithmatic Intensity, and synchronization patterns). CudaPerf operates in two stages: (1) an offline pairwise ranking module that learns to distinguish strong and weak program candidates via contrastive comparisons, and (2) an online RL training phase that jointly optimizes for correctness, performance, and structural efficiency through a unified reward signal. To further enhance learning, CudaPerf utilizes iterative refinement using execution feedback enabling progressive improvement of generated candidates. We also introduce a dataset comprising 2.9k C to CUDA and 1k PyTorch to CUDA programs, each paired with diverse input configurations and multiple CUDA implementations encompassing diverse optimization strategies. CudaPerf is evaluated across multiple benchmarks comprising both C to CUDA and PyTorch to CUDA transformations. Empirical findings suggest that CudaPerf significantly outperforms strong baselines, including Qwen-3-32B (for C to CUDA) and CUDA Agent (for PyTorch to CUDA) by achieving up to 5X & 3.32X improvements in speedup, and 17% & 7% improvements in correctness, respectively.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Reinforcement learning for reasoning, one of 17 directions in LLM reasoning. Builds on 23 of its 39 references that are in the map. First posted on arXiv 23 Jul 2026.

Builds on

23 of 39 references.

Closest ancestors · 3

Same direction, newest first

  1. 2025Reinforcement Learning for Reasoning in Large Language Models with One Training Example29 Apr
  2. 2025Synthetic Data Generation & Multi-Step RL for Reasoning & Tool Use7 Apr
  3. 2025SWE-RL: Advancing LLM Reasoning via Reinforcement Learning on Open Software Evolution25 Feb

Other references · 20

  1. 2021Training Verifiers to Solve Math Word Problemscited by 9,743
  2. 2021Evaluating Large Language Models Trained on Codecited by 8,053
  3. 2023QLoRA: Efficient Finetuning of Quantized LLMscited by 3,355
  4. 2023Code Llama: Open Foundation Models for Codecited by 2,105
  5. 2024Qwen2.5-Coder Technical Reportcited by 1,151
  6. 2022CodeGen: An Open Large Language Model for Code with Multi-Turn Program Synthesiscited by 876
  7. 2025RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learningcited by 357
  8. 2025KernelBench: Can LLMs Write Efficient GPU Kernels?cited by 206
  9. 2021CodeNet: A Large-Scale AI for Code Dataset for Learning a Diversity of Coding Taskscited by 199
  10. 2024RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learningcited by 146
  11. 2023Learning Performance-Improving Code Editscited by 103
  12. 2025Kevin: Multi-Turn RL for Generating CUDA Kernelscited by 73
8 more
  1. 2025CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learningcited by 59
  2. 2025Towards Robust Agentic CUDA Kernel Benchmarking, Verification, and Optimizationcited by 52
  3. 2026CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generationcited by 41
  4. 2025QiMeng-CodeV-R1: Reasoning-Enhanced Verilog Generationcited by 34
  5. 2024ECCO: Can We Improve Model-Generated Code Efficiency Without Sacrificing Functional Correctness?cited by 25
  6. 2024CodeRosetta: Pushing the Boundaries of Unsupervised Code Translation for Parallel Programmingcited by 19
  7. 2025AKG kernel Agent: A Multi-Agent Framework for Cross-Platform Kernel Synthesiscited by 7
  8. 2026OptiML: An End-to-End Framework for Program Synthesis and CUDA Kernel Optimizationcited by 3

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026LEAP: Lean Environment-Feedback via Adaptive Pruning for Code RL in GPU Kernel Generation
  2. 2026Reinforcement Learning for Code Optimization
  3. 2026Improving Small Language Models for Code Generation with Reinforcement Learning from Verification Feedback
  4. 2025DRIVE: Data Curation Best Practices for Reinforcement Learning with Verifiable Reward in Competitive Code Generation
  5. 2026CUDA-Harness: Harnessing Agentic CUDA Kernel Generation and Optimization from Natural Language
  6. 2026When Do Intrinsic Rewards Work for Code Reasoning? A Comprehensive Study
  7. 2025From Large to Small: Transferring CUDA Optimization Expertise via Reasoning Graph
  8. 2025CodeRL+: Improving Code Generation via Reinforcement with Execution Semantics Alignment