@inproceedings{liu-etal-2026-architecture,
title = "Architecture-Conditioned Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning",
author = "Liu, Kai and
Dong, Peijie and
Xinchen, Xie and
Gao, Jianfei and
Guo, Qipeng and
Chu, Xiaowen and
Zhang, Shaoting and
Chen, Kai",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.851/",
pages = "19713--19731",
ISBN = "979-8-89176-459-0",
abstract = "The rapid progress of reasoning and agentic large language models (LLMs) has increased the demand for long-context inference, but self-attention (SA) scales quadratically with context length. To address this, we propose SWARR (Sliding-Window Attention with Reinforced Adaptation for Math Reasoning), a practical recipe for adapting SWA models to mathematical reasoning. SWARR has two stages: (1) efficient conversion from a pretrained SA model to SWA with supervised fine-tuning (SFT), which avoids pretraining a new base model, and (2) architecture-aware policy adaptation with reinforcement learning (RL). We find that SWA still underperforms SA after SFT, and we hypothesize that this gap is caused in part by a data-architecture mismatch: most SFT data are prepared for SA models and may contain long-range dependencies that are difficult for SWA to model. Because RL is architecture-aware and optimizes self-generated trajectories under the SWA constraint, it can adapt training trajectories to better match the SWA architecture. Experiments on mathematical reasoning benchmarks show that this architecture-aware RL substantially narrows the gap between SWA and SA, recovering much of the accuracy lost during SWA conversion while preserving the efficiency benefits of linear-complexity attention."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="liu-etal-2026-architecture">
<titleInfo>
<title>Architecture-Conditioned Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kai</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Peijie</namePart>
<namePart type="family">Dong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xie</namePart>
<namePart type="family">Xinchen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jianfei</namePart>
<namePart type="family">Gao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qipeng</namePart>
<namePart type="family">Guo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaowen</namePart>
<namePart type="family">Chu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shaoting</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kai</namePart>
<namePart type="family">Chen</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>The rapid progress of reasoning and agentic large language models (LLMs) has increased the demand for long-context inference, but self-attention (SA) scales quadratically with context length. To address this, we propose SWARR (Sliding-Window Attention with Reinforced Adaptation for Math Reasoning), a practical recipe for adapting SWA models to mathematical reasoning. SWARR has two stages: (1) efficient conversion from a pretrained SA model to SWA with supervised fine-tuning (SFT), which avoids pretraining a new base model, and (2) architecture-aware policy adaptation with reinforcement learning (RL). We find that SWA still underperforms SA after SFT, and we hypothesize that this gap is caused in part by a data-architecture mismatch: most SFT data are prepared for SA models and may contain long-range dependencies that are difficult for SWA to model. Because RL is architecture-aware and optimizes self-generated trajectories under the SWA constraint, it can adapt training trajectories to better match the SWA architecture. Experiments on mathematical reasoning benchmarks show that this architecture-aware RL substantially narrows the gap between SWA and SA, recovering much of the accuracy lost during SWA conversion while preserving the efficiency benefits of linear-complexity attention.</abstract>
<identifier type="citekey">liu-etal-2026-architecture</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.851/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>19713</start>
<end>19731</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Architecture-Conditioned Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning
%A Liu, Kai
%A Dong, Peijie
%A Xinchen, Xie
%A Gao, Jianfei
%A Guo, Qipeng
%A Chu, Xiaowen
%A Zhang, Shaoting
%A Chen, Kai
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F liu-etal-2026-architecture
%X The rapid progress of reasoning and agentic large language models (LLMs) has increased the demand for long-context inference, but self-attention (SA) scales quadratically with context length. To address this, we propose SWARR (Sliding-Window Attention with Reinforced Adaptation for Math Reasoning), a practical recipe for adapting SWA models to mathematical reasoning. SWARR has two stages: (1) efficient conversion from a pretrained SA model to SWA with supervised fine-tuning (SFT), which avoids pretraining a new base model, and (2) architecture-aware policy adaptation with reinforcement learning (RL). We find that SWA still underperforms SA after SFT, and we hypothesize that this gap is caused in part by a data-architecture mismatch: most SFT data are prepared for SA models and may contain long-range dependencies that are difficult for SWA to model. Because RL is architecture-aware and optimizes self-generated trajectories under the SWA constraint, it can adapt training trajectories to better match the SWA architecture. Experiments on mathematical reasoning benchmarks show that this architecture-aware RL substantially narrows the gap between SWA and SA, recovering much of the accuracy lost during SWA conversion while preserving the efficiency benefits of linear-complexity attention.
%U https://aclanthology.org/2026.emnlp-main.851/
%P 19713-19731
Markdown (Informal)
[Architecture-Conditioned Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning](https://aclanthology.org/2026.emnlp-main.851/) (Liu et al., EMNLP 2026)
ACL
- Kai Liu, Peijie Dong, Xie Xinchen, Jianfei Gao, Qipeng Guo, Xiaowen Chu, Shaoting Zhang, and Kai Chen. 2026. Architecture-Conditioned Reinforcement Learning Makes Sliding-Window Attention Competitive in Math Reasoning. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 19713–19731, Budapest, Hungary. Association for Computational Linguistics.