mlsanityData to 14 Oct 2026

Kevin: Multi-Turn RL for Generating CUDA Kernels

arXiv 2507.11948 ↗First posted 16 Jul 2025cs.LG cs.AI cs.PFcited by 78

Not on the map. Why?

Cited by 25 in LLM agents and tool use, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

Writing GPU kernels is a challenging task and critical for AI systems' efficiency. It is also highly iterative: domain experts write code and improve performance through execution feedback. Moreover, it presents verifiable rewards like correctness and speedup, making it a natural environment to apply Reinforcement Learning (RL). To explicitly incorporate the iterative nature of this process into training, we develop a flexible multi-turn RL recipe that addresses unique challenges encountered in real-world settings, such as learning from long trajectories and effective reward attribution across turns. We present Kevin - K(ernel D)evin, the first model trained with multi-turn RL for CUDA kernel generation and optimization. In our evaluation setup, Kevin shows significant gains over its base model (QwQ-32B), improving correctness of generated kernels (in pure CUDA) from 56% to 82% and mean speedup from 0.53x to 1.10x of baseline (PyTorch Eager), and surpassing frontier models like o4-mini (0.78x). Finally, we study its behavior across test-time scaling axes: we found scaling serial refinement more beneficial than parallel sampling. In particular, when given more refinement turns, Kevin shows a higher rate of improvement.

Key facts. Cited by 78 papers in the map (data through 14 Oct 2026). Builds on 20 of its 43 references that are in the map. First posted on arXiv 16 Jul 2025.

Builds on

20 of 43 references.

References · 20

  1. 2017Proximal Policy Optimization Algorithmscited by 15,215
  2. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
  3. 2021Evaluating Large Language Models Trained on Codecited by 8,320
  4. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,330
  5. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,944
  6. 2022FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awarenesscited by 3,673
  7. 2023FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioningcited by 2,598
  8. 2024Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameterscited by 1,986
  9. 2016Concrete Problems in AI Safetycited by 1,642
  10. 2024Large Language Monkeys: Scaling Inference Compute with Repeated Samplingcited by 991
  11. 2025Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?cited by 982
  12. 2017Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithmcited by 734
8 more
  1. 2016RL2: Fast Reinforcement Learning via Slow Reinforcement Learningcited by 503
  2. 2025Reinforcement Learning for Reasoning in Large Language Models with One Training Examplecited by 295
  3. 2025FlashInfer: Efficient and Customizable Attention Engine for LLM Inference Servingcited by 258
  4. 2025Optimizing Test-Time Compute via Meta Reinforcement Fine-Tuningcited by 123
  5. 2025TritonBench: Benchmarking Large Language Model Capabilities for Generating Triton Operatorscited by 75
  6. 2023ZeRO++: Extremely Efficient Collective Communication for Giant Model Trainingcited by 52
  7. 2025LLMs are Greedy Agents: Effects of RL Fine-tuning on Decision-Making Abilitiescited by 41
  8. 2025Putting the Value Back in RL: Better Test-Time Scaling by Unifying LLM Reasoners With Verifierscited by 14

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 19; 2026: 59.

78 papers.

Sort:

By area:

  1. 2025Astra: A Multi-Agent System for GPU Kernel Performance Optimizationcited by 51
  2. 2026CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generationcited by 47
  3. 2025CudaForge: An Agent Framework with Hardware Feedback for CUDA Kernel Optimizationcited by 45
  4. 2025MultiKernelBench: A Multi-Platform Benchmark for Kernel Generationcited by 38
  5. 2026FlashInfer-Bench: Building the Virtuous Cycle for AI-driven LLM Systemscited by 25
  6. 2025TritonRL: Training LLMs to Think and Code Triton Without Cheatingcited by 24
  7. 2026K-Search: LLM Kernel Generation via Co-Evolving Intrinsic World Modelcited by 24
  8. 2025AccelOpt: A Self-Improving LLM Agentic System for AI Accelerator Kernel Optimizationcited by 23
  9. 2026Dr. Kernel: Reinforcement Learning Done Right for Triton Kernel Generationscited by 22
  10. 2025ConCuR: Conciseness Makes State-of-the-Art Kernel Generationcited by 13
  11. 2025CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learningcited by 13
  12. 2026KernelFoundry: Hardware-aware evolutionary GPU kernel optimizationcited by 13
38 more
  1. 2025SwizzlePerf: Hardware-Aware LLMs for GPU Kernel Performance Optimizationcited by 11
  2. 2025Scaling Generalist Data-Analytic Agentscited by 11
  3. 2026Towards Automated Kernel Generation in the Era of LLMscited by 11
  4. 2026KernelSkill: A Multi-Agent Framework for GPU Kernel Optimizationcited by 11
  5. 2025From Large to Small: Transferring CUDA Optimization Expertise via Reasoning Graphcited by 10
  6. 2026Hardening Agent Benchmarks with Adversarial Hacker-Fixer Loopscited by 9
  7. 2026Understanding the Challenges in Iterative Generative Optimization with LLMscited by 8
  8. 2025cuPilot: A Strategy-Coordinated Multi-agent Framework for CUDA Kernel Evolutioncited by 7
  9. 2025AKG kernel Agent: A Multi-Agent Framework for Cross-Platform Kernel Synthesiscited by 7
  10. 2026AscendCraft: Automatic Ascend NPU Kernel Generation via DSL-Guided Transcompilationcited by 7
  11. 2026StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement Learningcited by 7
  12. 2026Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimizationcited by 6
  13. 2025LANPO: Bootstrapping Language and Numerical Feedback for Reinforcement Learning in LLMscited by 5
  14. 2025KernelEvolve: Scaling Agentic Kernel Coding for Heterogeneous AI Accelerators at Metacited by 5
  15. 2026Improving Efficiency of GPU Kernel Optimization Agents using a Domain-Specific Language and Speed-of-Light Guidancecited by 5
  16. 2025Optimizing PyTorch Inference with LLM-Based Multi-Agent Systemscited by 4
  17. 2026MaxCode: A Max-Reward Reinforcement Learning Framework for Automated Code Optimizationcited by 4
  18. 2026Fine-Tuning GPT-5 for GPU Kernel Generationcited by 4
  19. 2026Making LLMs Optimize Multi-Scenario CUDA Kernels Like Expertscited by 4
  20. 2026InCoder-32B-Thinking: Industrial Code World Model for Thinkingcited by 4
  21. 2025PerfCoder: Large Language Models for Interpretable Code Performance Optimizationcited by 3
  22. 2026DICE: Diffusion Large Language Models Excel at Generating CUDA Kernelscited by 3
  23. 2026OptiML: An End-to-End Framework for Program Synthesis and CUDA Kernel Optimizationcited by 3
  24. 2026MultiCube-RAG for Multi-hop Question Answeringcited by 3
  25. 2026Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesiscited by 3
  26. 2026AscendOptimizer: Episodic Agent for Ascend NPU Operator Optimizationcited by 3
  27. 2026AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generationcited by 3
  28. 2026AHD Agent: Agentic Reinforcement Learning for Automatic Heuristic Designcited by 3
  29. 2026FastKernels: Benchmarking GPU Kernel Generation in Productioncited by 3
  30. 2026Tail-Likelihood Reinforcement Learningcited by 3
  31. 2025KernelBand: Steering LLM-based Kernel Optimization via Hardware-Aware Multi-Armed Banditscited by 2
  32. 2026CUCo: An Agentic Framework for Compute and Communication Co-designcited by 2
  33. 2026Xe-Forge: Multi-Stage LLM-Powered Kernel Optimization for Intel GPUcited by 2
  34. 2026MusaCoder: Native GPU Kernel Generation with Full-Stack Training on Moore Threads GPUcited by 2
  35. 2025Evolution of Kernels: Automated RISC-V Kernel Optimization with Large Language Modelscited by 1
  36. 2025HipKittens: Fast and Furious AMD Kernelscited by 1
  37. 2026TL-GRPO: Turn-Level RL for Reasoning-Guided Iterative Optimizationcited by 1
  38. 2026Surprisal-Guided Selection: Compute-Optimal Test-Time Strategies for Execution-Grounded Code Generationcited by 1

50 most cited of 78 All 78 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2025CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning
  2. 2026KernelBlaster: Continual Cross-Task CUDA Optimization via Memory-Augmented In-Context Reinforcement Learning
  3. 2026LEAP: Lean Environment-Feedback via Adaptive Pruning for Code RL in GPU Kernel Generation
  4. 2026Kernel Forge: An Agent Harness for LLM-based Generation and Optimization of CUDA Kernels
  5. 2025Towards Robust Agentic CUDA Kernel Benchmarking, Verification, and Optimization
  6. 2026CUDA-Harness: Harnessing Agentic CUDA Kernel Generation and Optimization from Natural Language
  7. 2026How to Train Your Little Dragon: Teaching a Small Language Model to Construct and Refine GPU Optimization Scripts
  8. 2026KernelBench-Verified: Do LLM-Generated Kernels Actually Beat PyTorch?