DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models Article Swipe

PDF

Shuming Shi , Wenjing Zhang , Jun Yan , Kai Wang , Zhaoxiang Liu , Shiguo Lian ·

YOU? · · 2025 · Open Access · · DOI: https://doi.org/10.48550/arxiv.2503.04472

Recent advancements in slow thinking reasoning models have shown exceptional performance in complex reasoning tasks. However, these models often exhibit overthinking (generating redundant reasoning steps for simple problems), leading to excessive computational resource usage. 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 the length of Chain-of-Thought (CoT) based on problem difficulty. We first propose a Token Length Budget (TLB) metric to quantify difficulty, then leverage budget-aware reward shaping and budget preference optimization to implement DAST. DAST penalizes overlong responses for simple tasks while incentivizing sufficient reasoning for complex problems. Experiments on diverse datasets and model scales demonstrate that DAST effectively mitigates overthinking (reducing token usage by over 30\% on average) while preserving reasoning accuracy on complex problems. Our codes and models are available at https://github.com/AnonymousUser0520/AnonymousRepo01.

Related Topics

Truth And Reconciliation Commission Of Canada

Reich Ministry Of Public Enlightenment And Propaganda

Rick Hurst

Fuck

Degenerate Art Exhibition

Concepts

No concepts available.

Metadata

Type: preprint
Language: en
Landing Page: http://arxiv.org/abs/2503.04472
PDF: https://arxiv.org/pdf/2503.04472
OA Status: green
OpenAlex ID: https://openalex.org/W4416113410

All OpenAlex metadata

Raw OpenAlex JSON

OpenAlex ID: https://openalex.org/W4416113410

Canonical identifier for this work in OpenAlex
DOI: https://doi.org/10.48550/arxiv.2503.04472

Digital Object Identifier
Title: DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models

Work title
Type: preprint

OpenAlex work type
Language: en

Primary language
Publication year: 2025

Year of publication
Publication date: 2025-03-06

Full publication date if available
Authors: Shuming Shi, Wenjing Zhang, Jun Yan, Kai Wang, Zhaoxiang Liu, Shiguo Lian

List of authors in order
Landing page: https://arxiv.org/abs/2503.04472

Publisher landing page
PDF URL: https://arxiv.org/pdf/2503.04472

Direct link to full text PDF
Open access: Yes

Whether a free full text is available
OA status: green

Open access status per OpenAlex
OA URL: https://arxiv.org/pdf/2503.04472

Direct OA link when available
Cited by: 0

Total citation count in OpenAlex

Full payload

id	https://openalex.org/W4416113410
doi	https://doi.org/10.48550/arxiv.2503.04472
ids.doi	https://doi.org/10.48550/arxiv.2503.04472
ids.openalex	https://openalex.org/W4416113410
fwci
type	preprint
title	DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models
biblio.issue
biblio.volume
biblio.last_page
biblio.first_page
is_xpac	False
apc_list
apc_paid
language	en
locations[0].id	pmh:oai:arXiv.org:2503.04472
locations[0].is_oa	True
locations[0].source.id	https://openalex.org/S4306400194
locations[0].source.issn
locations[0].source.type	repository
locations[0].source.is_oa	True
locations[0].source.issn_l
locations[0].source.is_core	False
locations[0].source.is_in_doaj	False
locations[0].source.display_name	arXiv (Cornell University)
locations[0].source.host_organization	https://openalex.org/I205783295
locations[0].source.host_organization_name	Cornell University
locations[0].source.host_organization_lineage	https://openalex.org/I205783295
locations[0].license
locations[0].pdf_url	https://arxiv.org/pdf/2503.04472
locations[0].version	submittedVersion
locations[0].raw_type	text
locations[0].license_id
locations[0].is_accepted	False
locations[0].is_published	False
locations[0].raw_source_name
locations[0].landing_page_url	http://arxiv.org/abs/2503.04472
locations[1].id	doi:10.48550/arxiv.2503.04472
locations[1].is_oa	True
locations[1].source.id	https://openalex.org/S4306400194
locations[1].source.issn
locations[1].source.type	repository
locations[1].source.is_oa	True
locations[1].source.issn_l
locations[1].source.is_core	False
locations[1].source.is_in_doaj	False
locations[1].source.display_name	arXiv (Cornell University)
locations[1].source.host_organization	https://openalex.org/I205783295
locations[1].source.host_organization_name	Cornell University
locations[1].source.host_organization_lineage	https://openalex.org/I205783295
locations[1].license
locations[1].pdf_url
locations[1].version
locations[1].raw_type	article
locations[1].license_id
locations[1].is_accepted	False
locations[1].is_published
locations[1].raw_source_name
locations[1].landing_page_url	https://doi.org/10.48550/arxiv.2503.04472
indexed_in	arxiv, datacite
authorships[0].author.id	https://openalex.org/A5087920747
authorships[0].author.orcid	https://orcid.org/0000-0001-7018-0682
authorships[0].author.display_name	Shuming Shi
authorships[0].author_position	middle
authorships[0].raw_author_name	Shi, Shuming
authorships[0].is_corresponding	False
authorships[1].author.id	https://openalex.org/A5100407210
authorships[1].author.orcid	https://orcid.org/0000-0002-6694-6072
authorships[1].author.display_name	Wenjing Zhang
authorships[1].author_position	middle
authorships[1].raw_author_name	Zhang, Wenjing
authorships[1].is_corresponding	False
authorships[2].author.id	https://openalex.org/A5108057313
authorships[2].author.orcid	https://orcid.org/0000-0002-1303-1952
authorships[2].author.display_name	Jun Yan
authorships[2].author_position	last
authorships[2].raw_author_name	Yan, Jiangze
authorships[2].is_corresponding	False
authorships[3].author.id	https://openalex.org/A5041062995
authorships[3].author.orcid	https://orcid.org/0000-0002-8813-1448
authorships[3].author.display_name	Kai Wang
authorships[3].author_position	middle
authorships[3].raw_author_name	Wang, Kai
authorships[3].is_corresponding	False
authorships[4].author.id	https://openalex.org/A5101742160
authorships[4].author.orcid	https://orcid.org/0000-0003-3404-6229
authorships[4].author.display_name	Zhaoxiang Liu
authorships[4].author_position	middle
authorships[4].raw_author_name	Liu, Zhaoxiang
authorships[4].is_corresponding	False
authorships[5].author.id	https://openalex.org/A5066958531
authorships[5].author.orcid	https://orcid.org/0000-0003-4308-7049
authorships[5].author.display_name	Shiguo Lian
authorships[5].author_position	middle
authorships[5].raw_author_name	Lian, Shiguo
authorships[5].is_corresponding	False
has_content.pdf	False
has_content.grobid_xml	False
is_paratext	False
open_access.is_oa	True
open_access.oa_url	https://arxiv.org/pdf/2503.04472
open_access.oa_status	green
open_access.any_repository_has_fulltext	False
created_date	2025-10-10T00:00:00
display_name	DAST: Difficulty-Adaptive Slow-Thinking for Large Reasoning Models
has_fulltext	False
is_retracted	False
updated_date	2025-11-28T05:25:44.843702
primary_topic
cited_by_count	0
locations_count	2
best_oa_location.id	pmh:oai:arXiv.org:2503.04472
best_oa_location.is_oa	True
best_oa_location.source.id	https://openalex.org/S4306400194
best_oa_location.source.issn
best_oa_location.source.type	repository
best_oa_location.source.is_oa	True
best_oa_location.source.issn_l
best_oa_location.source.is_core	False
best_oa_location.source.is_in_doaj	False
best_oa_location.source.display_name	arXiv (Cornell University)
best_oa_location.source.host_organization	https://openalex.org/I205783295
best_oa_location.source.host_organization_name	Cornell University
best_oa_location.source.host_organization_lineage	https://openalex.org/I205783295
best_oa_location.license
best_oa_location.pdf_url	https://arxiv.org/pdf/2503.04472
best_oa_location.version	submittedVersion
best_oa_location.raw_type	text
best_oa_location.license_id
best_oa_location.is_accepted	False
best_oa_location.is_published	False
best_oa_location.raw_source_name
best_oa_location.landing_page_url	http://arxiv.org/abs/2503.04472
primary_location.id	pmh:oai:arXiv.org:2503.04472
primary_location.is_oa	True
primary_location.source.id	https://openalex.org/S4306400194
primary_location.source.issn
primary_location.source.type	repository
primary_location.source.is_oa	True
primary_location.source.issn_l
primary_location.source.is_core	False
primary_location.source.is_in_doaj	False
primary_location.source.display_name	arXiv (Cornell University)
primary_location.source.host_organization	https://openalex.org/I205783295
primary_location.source.host_organization_name	Cornell University
primary_location.source.host_organization_lineage	https://openalex.org/I205783295
primary_location.license
primary_location.pdf_url	https://arxiv.org/pdf/2503.04472
primary_location.version	submittedVersion
primary_location.raw_type	text
primary_location.license_id
primary_location.is_accepted	False
primary_location.is_published	False
primary_location.raw_source_name
primary_location.landing_page_url	http://arxiv.org/abs/2503.04472
publication_date	2025-03-06
publication_year	2025
referenced_works_count	0
abstract_inverted_index.a	60, 81
abstract_inverted_index.We	78
abstract_inverted_index.at	150
abstract_inverted_index.by	132
abstract_inverted_index.in	2, 11
abstract_inverted_index.of	71
abstract_inverted_index.on	46, 75, 117, 135, 141
abstract_inverted_index.to	29, 66, 87, 99
abstract_inverted_index.Our	144
abstract_inverted_index.and	95, 120, 146
abstract_inverted_index.are	148
abstract_inverted_index.for	25, 106, 113
abstract_inverted_index.the	69
abstract_inverted_index.30\%	134
abstract_inverted_index.DAST	102, 125
abstract_inverted_index.Slow	57
abstract_inverted_index.This	53
abstract_inverted_index.have	7
abstract_inverted_index.over	133
abstract_inverted_index.risk	43
abstract_inverted_index.slow	3
abstract_inverted_index.that	49, 63, 124
abstract_inverted_index.then	90
abstract_inverted_index.they	42
abstract_inverted_index.(CoT)	73
abstract_inverted_index.(TLB)	85
abstract_inverted_index.DAST.	101
abstract_inverted_index.Token	82
abstract_inverted_index.While	34
abstract_inverted_index.based	74
abstract_inverted_index.codes	145
abstract_inverted_index.first	79
abstract_inverted_index.model	121
abstract_inverted_index.novel	61
abstract_inverted_index.often	18
abstract_inverted_index.paper	54
abstract_inverted_index.shown	8
abstract_inverted_index.steps	24
abstract_inverted_index.tasks	48, 108
abstract_inverted_index.these	16
abstract_inverted_index.token	130
abstract_inverted_index.usage	131
abstract_inverted_index.while	109, 137
abstract_inverted_index.Budget	84
abstract_inverted_index.Length	83
abstract_inverted_index.Recent	0
abstract_inverted_index.adjust	68
abstract_inverted_index.budget	96
abstract_inverted_index.length	70
abstract_inverted_index.metric	86
abstract_inverted_index.models	6, 17, 65, 147
abstract_inverted_index.reduce	39
abstract_inverted_index.reward	93
abstract_inverted_index.scales	122
abstract_inverted_index.simple	26, 107
abstract_inverted_index.tasks.	14
abstract_inverted_index.usage.	33
abstract_inverted_index.(DAST),	59
abstract_inverted_index.complex	12, 114, 142
abstract_inverted_index.current	35
abstract_inverted_index.diverse	118
abstract_inverted_index.enables	64
abstract_inverted_index.exhibit	19
abstract_inverted_index.leading	28
abstract_inverted_index.problem	76
abstract_inverted_index.propose	80
abstract_inverted_index.require	50
abstract_inverted_index.shaping	94
abstract_inverted_index.tokens,	41
abstract_inverted_index.However,	15
abstract_inverted_index.Thinking	58
abstract_inverted_index.accuracy	140
abstract_inverted_index.average)	136
abstract_inverted_index.datasets	119
abstract_inverted_index.extended	51
abstract_inverted_index.leverage	91
abstract_inverted_index.overlong	104
abstract_inverted_index.quantify	88
abstract_inverted_index.resource	32
abstract_inverted_index.thinking	4
abstract_inverted_index.(reducing	129
abstract_inverted_index.available	149
abstract_inverted_index.degrading	44
abstract_inverted_index.excessive	30
abstract_inverted_index.framework	62
abstract_inverted_index.implement	100
abstract_inverted_index.mitigates	127
abstract_inverted_index.penalizes	103
abstract_inverted_index.problems.	115, 143
abstract_inverted_index.reasoning	5, 13, 23, 40, 112, 139
abstract_inverted_index.redundant	22
abstract_inverted_index.responses	105
abstract_inverted_index.uniformly	38
abstract_inverted_index.introduces	55
abstract_inverted_index.mitigation	36
abstract_inverted_index.preference	97
abstract_inverted_index.preserving	138
abstract_inverted_index.problems),	27
abstract_inverted_index.reasoning.	52
abstract_inverted_index.strategies	37
abstract_inverted_index.sufficient	111
abstract_inverted_index.(generating	21
abstract_inverted_index.Experiments	116
abstract_inverted_index.challenging	47
abstract_inverted_index.demonstrate	123
abstract_inverted_index.difficulty,	89
abstract_inverted_index.difficulty.	77
abstract_inverted_index.effectively	126
abstract_inverted_index.exceptional	9
abstract_inverted_index.performance	10, 45
abstract_inverted_index.advancements	1
abstract_inverted_index.autonomously	67
abstract_inverted_index.budget-aware	92
abstract_inverted_index.optimization	98
abstract_inverted_index.overthinking	20, 128
abstract_inverted_index.computational	31
abstract_inverted_index.incentivizing	110
abstract_inverted_index.Chain-of-Thought	72
abstract_inverted_index.Difficulty-Adaptive	56
abstract_inverted_index.https://github.com/AnonymousUser0520/AnonymousRepo01.	151
cited_by_percentile_year
countries_distinct_count	0
institutions_distinct_count	6
citation_normalized_percentile