REVIEW 16 cited by
Kevin: Multi-Turn RL for Generating CUDA Kernels
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Kevin: Multi-Turn RL for Generating CUDA Kernels
read the original abstract
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.
Forward citations
Cited by 16 Pith papers
-
FastKernels: Benchmarking GPU Kernel Generation in Production
FastKernels is a production-aligned benchmark covering 96.2% of HuggingFace Transformers that reveals state-of-the-art kernel agents deliver at most 0.94x aggregate speedup.
-
PassNet: Scaling Large Language Models for Graph Compiler Pass Generation
PassNet provides a dataset of 18K graphs and PassBench for LLM-generated compiler passes, with fine-tuned models achieving 2.67x gains on long-tail tasks where TorchInductor underperforms.
-
Step-TP: A Grounded, Step-Level Dataset with Chain-of-Thought Reasoning for LLM-Guided Tensor Program Optimization
Step-TP is a dataset providing grounded, atomic step-level IR transitions and CoT supervision to enable reliable multi-step LLM-guided tensor program optimization instead of end-to-end imitation.
-
CUDAHercules: Benchmarking Hardware-Aware Expert-level CUDA Optimization for LLMs
CUDAHercules benchmark demonstrates that leading LLMs generate functional CUDA code but fail to recover expert-level optimization strategies needed for peak performance on Ampere, Hopper, and Blackwell GPUs.
-
CUDABeaver: Benchmarking LLM-Based Automated CUDA Debugging
CUDABeaver shows LLM CUDA debuggers often degenerate code for test-passing at the cost of speed, with protocol-aware metrics shifting success rates by up to 40 percentage points.
-
Hawk: Harnessing Hardware-Aware Knowledge for High-Performance NPU Kernel Generation
Hawk raises NPU kernel generation accuracy from 49.4% to 80% and yields up to 2.2× speedups by retrieving and distilling structured hardware-aware knowledge without any model training.
-
Hawk: Harnessing Hardware-Aware Knowledge for High-Performance NPU Kernel Generation
Hawk is a training-free framework that boosts NPU kernel generation accuracy to 80% and achieves up to 2.2x speedup via hardware-aware knowledge synthesis, 2D retrieval, and effect-driven distillation.
-
SpecGen: Accelerating Agentic Kernel Optimization with Speculative Generation
SpecGen introduces speculative generation to fork non-reasoning kernel candidates during LLM reasoning traces, enabling early termination and parallel profiling to reduce end-to-end optimization time on H200 GPUs.
-
Learning When to Optimize: Verified Optimization Skills from Expert GPU-Kernel Lineages
KLineage derives verified optimization skills from backward lineages of expert GPU kernels to guide LLM agents toward higher-quality and more efficient kernels than memory-based baselines.
-
CUDABeaver: Benchmarking LLM-Based Automated CUDA Debugging
CUDABEAVER benchmark and pass@k(M,C,A) metric show LLM CUDA debugging success drops by up to 40 percentage points under strict performance requirements.
-
Kernel Foundry: A Diagnosis-driven Evolutionary Kernel Optimizer with Multi-Experts
Kernel Foundry improves GPU kernel correctness and performance on KernelBench via expert-guided evolutionary search with diagnostic feedback and an experience library, reaching 100% correctness on Level 2.
-
AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation
AdaExplore improves correctness and speed of Triton kernel generation by converting recurring failures into a memory of rules and organizing search as a tree that mixes local refinements with larger regenerations, yie...
-
InCoder-32B-Thinking: Industrial Code World Model for Thinking
InCoder-32B-Thinking uses error-feedback synthesized thinking traces and a code world model to reach top open-source scores on general and industrial code benchmarks including 81.3% on LiveCodeBench and 84.0% on CAD-Coder.
-
Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization
Kernel-Smith combines evolutionary search with RL post-training to generate optimized GPU kernels, achieving SOTA speedups on KernelBench that beat Gemini-3.0-pro and Claude-4.6-opus on NVIDIA Triton and generalize to...
-
The FIL Hypothesis: Inductive Biases Help with Kernel Engineering
The FIL Hypothesis claims that inductive biases outperform purely data-driven methods on GPU programming tasks with non-trivial feedback loops.
-
AscendOptimizer: Episodic Agent for Ascend NPU Operator Optimization
AscendOptimizer combines kernel rewinding for reusable experience with evolutionary search on hardware feedback to optimize Ascend NPU operators, delivering 1.21x geometric-mean speedup and faster performance on 53.47...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.