Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesis

Yujie Zheng, Zhuo Li, Shengtao Zhang, Jiaqian Wang, Junjie Sheng, Junchi Yan, Weinan Zhang, Ying Wen, Bo Tang, Muning Wen
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:163313-163339, 2026.

Abstract

Deploying Large Language Models to data-scarce programming domains poses significant challenges, particularly for kernel synthesis on emerging Domain-Specific Architectures where a “Data Wall” limits available training data. While models excel on data-rich platforms like CUDA, they suffer catastrophic performance drops on data-scarce ecosystems such as NPU programming. To overcome this cold-start barrier without expensive fine-tuning, we introduce EvoKernel, a self-evolving agentic framework that automates the lifecycle of kernel synthesis from initial drafting to continual refining. EvoKernel addresses this by formulating the synthesis process as a memory-based reinforcement learning task. Through a novel value-driven retrieval mechanism, it learns stage-specific Q-values that prioritize experiences based on their contribution to the current objective—whether bootstrapping a feasible draft or iteratively refining latency. Furthermore, by enabling cross-task memory sharing, the agent generalizes insights from simple to complex operators. By building an NPU variant of KernelBench and evaluating on it, EvoKernel improves frontier models’ correctness from 11.0% to 83.0% and achieves a median speedup of 3.60$\times$ over initial drafts through iterative refinement. This demonstrates that value-driven experience accumulation allows general-purpose models to master the kernel synthesis task on niche hardware ecosystems.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-zheng26h, title = {Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to {NPU} Kernel Synthesis}, author = {Zheng, Yujie and Li, Zhuo and Zhang, Shengtao and Wang, Jiaqian and Sheng, Junjie and Yan, Junchi and Zhang, Weinan and Wen, Ying and Tang, Bo and Wen, Muning}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {163313--163339}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/zheng26h/zheng26h.pdf}, url = {https://proceedings.mlr.press/v306/zheng26h.html}, abstract = {Deploying Large Language Models to data-scarce programming domains poses significant challenges, particularly for kernel synthesis on emerging Domain-Specific Architectures where a “Data Wall” limits available training data. While models excel on data-rich platforms like CUDA, they suffer catastrophic performance drops on data-scarce ecosystems such as NPU programming. To overcome this cold-start barrier without expensive fine-tuning, we introduce EvoKernel, a self-evolving agentic framework that automates the lifecycle of kernel synthesis from initial drafting to continual refining. EvoKernel addresses this by formulating the synthesis process as a memory-based reinforcement learning task. Through a novel value-driven retrieval mechanism, it learns stage-specific Q-values that prioritize experiences based on their contribution to the current objective—whether bootstrapping a feasible draft or iteratively refining latency. Furthermore, by enabling cross-task memory sharing, the agent generalizes insights from simple to complex operators. By building an NPU variant of KernelBench and evaluating on it, EvoKernel improves frontier models’ correctness from 11.0% to 83.0% and achieves a median speedup of 3.60$\times$ over initial drafts through iterative refinement. This demonstrates that value-driven experience accumulation allows general-purpose models to master the kernel synthesis task on niche hardware ecosystems.} }
Endnote
%0 Conference Paper %T Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesis %A Yujie Zheng %A Zhuo Li %A Shengtao Zhang %A Jiaqian Wang %A Junjie Sheng %A Junchi Yan %A Weinan Zhang %A Ying Wen %A Bo Tang %A Muning Wen %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-zheng26h %I PMLR %P 163313--163339 %U https://proceedings.mlr.press/v306/zheng26h.html %V 306 %X Deploying Large Language Models to data-scarce programming domains poses significant challenges, particularly for kernel synthesis on emerging Domain-Specific Architectures where a “Data Wall” limits available training data. While models excel on data-rich platforms like CUDA, they suffer catastrophic performance drops on data-scarce ecosystems such as NPU programming. To overcome this cold-start barrier without expensive fine-tuning, we introduce EvoKernel, a self-evolving agentic framework that automates the lifecycle of kernel synthesis from initial drafting to continual refining. EvoKernel addresses this by formulating the synthesis process as a memory-based reinforcement learning task. Through a novel value-driven retrieval mechanism, it learns stage-specific Q-values that prioritize experiences based on their contribution to the current objective—whether bootstrapping a feasible draft or iteratively refining latency. Furthermore, by enabling cross-task memory sharing, the agent generalizes insights from simple to complex operators. By building an NPU variant of KernelBench and evaluating on it, EvoKernel improves frontier models’ correctness from 11.0% to 83.0% and achieves a median speedup of 3.60$\times$ over initial drafts through iterative refinement. This demonstrates that value-driven experience accumulation allows general-purpose models to master the kernel synthesis task on niche hardware ecosystems.
APA
Zheng, Y., Li, Z., Zhang, S., Wang, J., Sheng, J., Yan, J., Zhang, W., Wen, Y., Tang, B. & Wen, M.. (2026). Towards Cold-Start Drafting and Continual Refining: A Value-Driven Memory Approach with Application to NPU Kernel Synthesis. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:163313-163339 Available from https://proceedings.mlr.press/v306/zheng26h.html.

Related Material