@inproceedings{chen-etal-2026-tokens,
title = "From Tokens to Regions: {CUDA}-Sensitive Instruction Tuning for {GPU} Kernel Generation",
author = "Chen, Wentao and
Zhu, Jiace and
Zeng, Qu and
Xiao, Qiaoling and
Duan, Liucheng and
Zou, An",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.2045/",
doi = "10.18653/2026.findings-emnlp.2045",
pages = "43505--43519",
ISBN = "979-8-89176-462-0",
abstract = "High-performance CUDA kernels are essential for scalable AI systems, while Large Language Models (LLMs) still struggle to generate correct kernels due to strict and implicit execution constraints. Existing LLM-based approaches either rely on costly agentic or reinforcement-learning (RL) pipelines, or adopt supervised fine-tuning (SFT) objectives that fail to explicitly model CUDA sensitivity, namely code tokens or regions tightly coupled with execution constraints. In this work, we investigate CUDA sensitivity from the perspective of token confidence patterns, showing that CUDA sensitivity appears at both token and region levels, where most CUDA-sensitive tokens are predicted with high confidence, while a smaller low-confidence subset forms regions corresponding to execution-critical structures. These findings suggest that effective CUDA kernel generation should both leverage high-confidence CUDA-sensitive tokens and preserve low-confidence CUDA-sensitive regions. Building on these insights, we propose \textbf{\underline{CU}DA-\underline{Se}nsitive Instruction \underline{T}uning (CuSeT)}, a low-cost post-training method within a simple SFT framework. CuSeT follows the principle of ``from tokens to regions'' by combining \textit{adaptive token-level masking} with \textit{region-aware sample reweighting}. Experiments show that CuSeT consistently improves both functional correctness and kernel performance across multiple model families and scales, outperforming standard SFT and advanced SFT variants, while achieving competitive performance against frontier CUDA kernel generation models with substantially lower inference cost."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="chen-etal-2026-tokens">
<titleInfo>
<title>From Tokens to Regions: CUDA-Sensitive Instruction Tuning for GPU Kernel Generation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Wentao</namePart>
<namePart type="family">Chen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiace</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qu</namePart>
<namePart type="family">Zeng</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qiaoling</namePart>
<namePart type="family">Xiao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Liucheng</namePart>
<namePart type="family">Duan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">An</namePart>
<namePart type="family">Zou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>High-performance CUDA kernels are essential for scalable AI systems, while Large Language Models (LLMs) still struggle to generate correct kernels due to strict and implicit execution constraints. Existing LLM-based approaches either rely on costly agentic or reinforcement-learning (RL) pipelines, or adopt supervised fine-tuning (SFT) objectives that fail to explicitly model CUDA sensitivity, namely code tokens or regions tightly coupled with execution constraints. In this work, we investigate CUDA sensitivity from the perspective of token confidence patterns, showing that CUDA sensitivity appears at both token and region levels, where most CUDA-sensitive tokens are predicted with high confidence, while a smaller low-confidence subset forms regions corresponding to execution-critical structures. These findings suggest that effective CUDA kernel generation should both leverage high-confidence CUDA-sensitive tokens and preserve low-confidence CUDA-sensitive regions. Building on these insights, we propose \underlineCUDA-\underlineSensitive Instruction \underlineTuning (CuSeT), a low-cost post-training method within a simple SFT framework. CuSeT follows the principle of “from tokens to regions” by combining adaptive token-level masking with region-aware sample reweighting. Experiments show that CuSeT consistently improves both functional correctness and kernel performance across multiple model families and scales, outperforming standard SFT and advanced SFT variants, while achieving competitive performance against frontier CUDA kernel generation models with substantially lower inference cost.</abstract>
<identifier type="citekey">chen-etal-2026-tokens</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.2045</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2045/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>43505</start>
<end>43519</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T From Tokens to Regions: CUDA-Sensitive Instruction Tuning for GPU Kernel Generation
%A Chen, Wentao
%A Zhu, Jiace
%A Zeng, Qu
%A Xiao, Qiaoling
%A Duan, Liucheng
%A Zou, An
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F chen-etal-2026-tokens
%X High-performance CUDA kernels are essential for scalable AI systems, while Large Language Models (LLMs) still struggle to generate correct kernels due to strict and implicit execution constraints. Existing LLM-based approaches either rely on costly agentic or reinforcement-learning (RL) pipelines, or adopt supervised fine-tuning (SFT) objectives that fail to explicitly model CUDA sensitivity, namely code tokens or regions tightly coupled with execution constraints. In this work, we investigate CUDA sensitivity from the perspective of token confidence patterns, showing that CUDA sensitivity appears at both token and region levels, where most CUDA-sensitive tokens are predicted with high confidence, while a smaller low-confidence subset forms regions corresponding to execution-critical structures. These findings suggest that effective CUDA kernel generation should both leverage high-confidence CUDA-sensitive tokens and preserve low-confidence CUDA-sensitive regions. Building on these insights, we propose \underlineCUDA-\underlineSensitive Instruction \underlineTuning (CuSeT), a low-cost post-training method within a simple SFT framework. CuSeT follows the principle of “from tokens to regions” by combining adaptive token-level masking with region-aware sample reweighting. Experiments show that CuSeT consistently improves both functional correctness and kernel performance across multiple model families and scales, outperforming standard SFT and advanced SFT variants, while achieving competitive performance against frontier CUDA kernel generation models with substantially lower inference cost.
%R 10.18653/2026.findings-emnlp.2045
%U https://aclanthology.org/2026.findings-emnlp.2045/
%U https://doi.org/10.18653/2026.findings-emnlp.2045
%P 43505-43519
Markdown (Informal)
[From Tokens to Regions: CUDA-Sensitive Instruction Tuning for GPU Kernel Generation](https://aclanthology.org/2026.findings-emnlp.2045/) (Chen et al., Findings 2026)
ACL