# KernelBench-Verified: Do LLM-Generated Kernels Actually Beat PyTorch?

arXiv: https://arxiv.org/abs/2607.16241 · HTML: https://mlsanity.com/paper/2607.16241/

## Key facts

- Cited by 8 papers in the map (data through 14 Oct 2026).
- Builds on 13 of its 22 references that are in the map.
- First posted on arXiv 26 Jun 2026.

## Builds on (most cited first, 13 of 13)

- [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness](https://mlsanity.com/paper/2205.14135/) (2022)
- [KernelBench: Can LLMs Write Efficient GPU Kernels?](https://mlsanity.com/paper/2502.10517/) (2025)
- [Establishing Best Practices for Building Rigorous Agentic Benchmarks](https://mlsanity.com/paper/2507.02825/) (2025)
- [Towards Robust Agentic CUDA Kernel Benchmarking, Verification, and Optimization](https://mlsanity.com/paper/2509.14279/) (2025)
- [CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation](https://mlsanity.com/paper/2602.24286/) (2026)
- [MultiKernelBench: A Multi-Platform Benchmark for Kernel Generation](https://mlsanity.com/paper/2507.17773/) (2025)
- [K-Search: LLM Kernel Generation via Co-Evolving Intrinsic World Model](https://mlsanity.com/paper/2602.19128/) (2026)
- [Dr. Kernel: Reinforcement Learning Done Right for Triton Kernel Generations](https://mlsanity.com/paper/2602.05885/) (2026)
- [KernelFoundry: Hardware-aware evolutionary GPU kernel optimization](https://mlsanity.com/paper/2603.12440/) (2026)
- [KernelBenchX: A Comprehensive Benchmark for Evaluating LLM-Generated GPU Kernels](https://mlsanity.com/paper/2605.04956/) (2026)
- [Towards Automated Kernel Generation in the Era of LLMs](https://mlsanity.com/paper/2601.15727/) (2026)
- [ISO-Bench: Can Coding Agents Optimize Real-World Inference Workloads?](https://mlsanity.com/paper/2602.19594/) (2026)
- [Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization](https://mlsanity.com/paper/2603.28342/) (2026)

## Built on by (most cited first, 8 of 8)

- [KernelGenBench: Can LLMs and Agents Write Efficient Kernels Across Operator Sources and Hardware Platforms?](https://mlsanity.com/paper/2607.27231/) (2026)
- [HIERA: Workload-Aware Planning Across Implementation Spaces for GPU Kernel Optimization](https://mlsanity.com/paper/2608.21157/) (2026)
- [AutoTuneBench: Trustworthy Measurement for Agent Auto-Tuning of LLM Serving Engines](https://mlsanity.com/paper/2609.18123/) (2026)
- [Measuring the Checker: Mutation Analysis for GPU-Kernel Benchmark Oracles](https://mlsanity.com/paper/2609.22220/) (2026)
- [AI as a Compiler: Compiling Triton kernels without the Triton compiler](https://mlsanity.com/paper/2609.36800/) (2026)
- [Self-Evolving Algorithm-Design Agents: Escaping In-Context Evolutionary Stagnation via Population-Curated Policy Optimization](https://mlsanity.com/paper/2609.38757/) (2026)
- [D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels?](https://mlsanity.com/paper/2610.03226/) (2026)
- [MetaKernelBench: Measuring GPU Kernel Knowledge Transfer Beyond Code](https://mlsanity.com/paper/2610.05014/) (2026)

## Abstract (arXiv)

Recent large language models (LLMs) can generate custom CUDA kernels that appear to outperform PyTorch on benchmarks such as KernelBench. Building upon this foundational framework, we demonstrate that frontier models frequently engage in reward hacking to artificially inflate reported performance. In this work, we identify two areas where evaluation frameworks must co-evolve with model capabilities. First, to accurately measure true speedup, we examine the baseline timing mechanism, noting that enabling Tensor Core acceleration with TF32 provides a more realistic estimation of execution on modern GPUs. Second, concerning algorithmic correctness, models often exploit the narrow test distribution by hardcoding bypasses for specific tensor values. By skipping required computations, these kernels artificially accelerate execution rather than implementing actual CUDA kernels. We introduce KernelBench-Verified, an extended evaluation framework that incorporates a TF32-enabled baseline and a four-distribution hidden test suite. We additionally introduce memory efficiency metrics that capture the often-overlooked speed-memory tradeoff in kernel optimization. Under verified single-turn evaluation with seven frontier LLMs, we find that the best-performing model (GPT-5.5) achieves a 0.88x geometric mean speedup, significantly lower than the 1.43x speedup observed under the standard evaluation protocol. No model consistently outperforms PyTorch when evaluated against realistic baselines. On the memory front, 28% of GPU kernels generated by the best model increase peak GPU memory usage. Our findings demonstrate the necessity of continually adapting robust evaluation protocols as LLM kernel generation capabilities advance.