@inproceedings{shen-etal-2025-dast,
title = "{DAST}: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models",
author = "Shen, Yi and
Zhang, Jian and
Huang, Jieyun and
Shi, Shuming and
Zhang, Wenjing and
Yan, Jiangze and
Wang, Ning and
Wang, Kai and
Liu, Zhaoxiang and
Lian, Shiguo",
editor = "Potdar, Saloni and
Rojas-Barahona, Lina and
Montella, Sebastien",
booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track",
month = nov,
year = "2025",
address = "Suzhou (China)",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.emnlp-industry.160/",
pages = "2322--2331",
ISBN = "979-8-89176-333-3",
abstract = "Recent advancements in slow-thinking reasoning models have shown exceptional performance in complex reasoning tasks. However, their tendency for ``overthinking'' on simple problems leads to excessive computational resource usage and increased inference latency, which hinders their widespread industrial adoption. While current mitigation strategies uniformly reduce reasoning tokens, they risk degrading performance on challenging tasks that require extended reasoning. This paper introduces Difficulty-Adaptive Slow-Thinking (DAST), a novel framework that enables models to autonomously adjust Chain-of-Thought (CoT) length based on problem difficulty. We propose a Token Length Budget (TLB) metric and leverage budget-aware preference optimization to implement DAST, which penalizes inefficiency on simple problems while incentivizing deep reasoning for complex ones. Experiments demonstrate DAST{'}s significant value for practical application: it effectively mitigates overthinking, substantially lowering costs and latency{---}while crucially preserving high accuracy on complex problems, paving the way for the efficient application of advanced reasoning models."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shen-etal-2025-dast">
<titleInfo>
<title>DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yi</namePart>
<namePart type="family">Shen</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jian</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jieyun</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shuming</namePart>
<namePart type="family">Shi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenjing</namePart>
<namePart type="family">Zhang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jiangze</namePart>
<namePart type="family">Yan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ning</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kai</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zhaoxiang</namePart>
<namePart type="family">Liu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shiguo</namePart>
<namePart type="family">Lian</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track</title>
</titleInfo>
<name type="personal">
<namePart type="given">Saloni</namePart>
<namePart type="family">Potdar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lina</namePart>
<namePart type="family">Rojas-Barahona</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sebastien</namePart>
<namePart type="family">Montella</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Suzhou (China)</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-333-3</identifier>
</relatedItem>
<abstract>Recent advancements in slow-thinking reasoning models have shown exceptional performance in complex reasoning tasks. However, their tendency for “overthinking” on simple problems leads to excessive computational resource usage and increased inference latency, which hinders their widespread industrial adoption. While current mitigation strategies uniformly reduce reasoning tokens, they risk degrading performance on challenging tasks that require extended reasoning. This paper introduces Difficulty-Adaptive Slow-Thinking (DAST), a novel framework that enables models to autonomously adjust Chain-of-Thought (CoT) length based on problem difficulty. We propose a Token Length Budget (TLB) metric and leverage budget-aware preference optimization to implement DAST, which penalizes inefficiency on simple problems while incentivizing deep reasoning for complex ones. Experiments demonstrate DAST’s significant value for practical application: it effectively mitigates overthinking, substantially lowering costs and latency—while crucially preserving high accuracy on complex problems, paving the way for the efficient application of advanced reasoning models.</abstract>
<identifier type="citekey">shen-etal-2025-dast</identifier>
<location>
<url>https://aclanthology.org/2025.emnlp-industry.160/</url>
</location>
<part>
<date>2025-11</date>
<extent unit="page">
<start>2322</start>
<end>2331</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models
%A Shen, Yi
%A Zhang, Jian
%A Huang, Jieyun
%A Shi, Shuming
%A Zhang, Wenjing
%A Yan, Jiangze
%A Wang, Ning
%A Wang, Kai
%A Liu, Zhaoxiang
%A Lian, Shiguo
%Y Potdar, Saloni
%Y Rojas-Barahona, Lina
%Y Montella, Sebastien
%S Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track
%D 2025
%8 November
%I Association for Computational Linguistics
%C Suzhou (China)
%@ 979-8-89176-333-3
%F shen-etal-2025-dast
%X Recent advancements in slow-thinking reasoning models have shown exceptional performance in complex reasoning tasks. However, their tendency for “overthinking” on simple problems leads to excessive computational resource usage and increased inference latency, which hinders their widespread industrial adoption. While current mitigation strategies uniformly reduce reasoning tokens, they risk degrading performance on challenging tasks that require extended reasoning. This paper introduces Difficulty-Adaptive Slow-Thinking (DAST), a novel framework that enables models to autonomously adjust Chain-of-Thought (CoT) length based on problem difficulty. We propose a Token Length Budget (TLB) metric and leverage budget-aware preference optimization to implement DAST, which penalizes inefficiency on simple problems while incentivizing deep reasoning for complex ones. Experiments demonstrate DAST’s significant value for practical application: it effectively mitigates overthinking, substantially lowering costs and latency—while crucially preserving high accuracy on complex problems, paving the way for the efficient application of advanced reasoning models.
%U https://aclanthology.org/2025.emnlp-industry.160/
%P 2322-2331
Markdown (Informal)
[DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models](https://aclanthology.org/2025.emnlp-industry.160/) (Shen et al., EMNLP 2025)
ACL
- Yi Shen, Jian Zhang, Jieyun Huang, Shuming Shi, Wenjing Zhang, Jiangze Yan, Ning Wang, Kai Wang, Zhaoxiang Liu, and Shiguo Lian. 2025. DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 2322–2331, Suzhou (China). Association for Computational Linguistics.