DeepResearchGuard: Deep Research with Open-Domain Evaluation and Multi-Stage Guardrails for Safety Article Swipe

PDF

Henry Peng Zou , Dongyuan Li , Andrea Zangari , Jing Guo , Chunyan Miao , Liang He , Renhe Jiang , Philip S. Yu ·

YOU? · · 2025 · Open Access · · DOI: https://doi.org/10.48550/arxiv.2510.10994

Deep research frameworks have shown promising capabilities in synthesizing comprehensive reports from web sources. While deep research possesses significant potential to address complex issues through planning and research cycles, existing frameworks are deficient in sufficient evaluation procedures and stage-specific protections. They typically treat evaluation as exact match accuracy of question-answering, but overlook crucial aspects of report quality such as credibility, coherence, breadth, depth, and safety. This oversight may result in hazardous or malicious sources being integrated into the final report. To address these issues, we introduce DEEPRESEARCHGUARD, a comprehensive framework featuring four-stage safeguards with open-domain evaluation of references and reports. We assess performance across multiple metrics, e.g., defense success rate and over-refusal rate, and five key report dimensions. In the absence of a suitable safety benchmark, we introduce DRSAFEBENCH, a stage-wise benchmark for deep research safety. Our evaluation spans diverse state-of-the-art LLMs, including GPT-4o, Gemini-2.5-flash, DeepSeek-v3, and o4-mini. DEEPRESEARCHGUARD achieves an average defense success rate improvement of 18.16% while reducing over-refusal rate by 6%. The input guard provides the most substantial early-stage protection by filtering out obvious risks, while the plan and research guards enhance citation discipline and source credibility. Through extensive experiments, we show that DEEPRESEARCHGUARD enables comprehensive open-domain evaluation and stage-aware defenses that effectively block harmful content propagation, while systematically improving report quality without excessive over-refusal rates. The code can be found via https://github.com/Jasonya/DeepResearchGuard.

Related Topics

Isfahan Nuclear Technology/Research Center

Concepts

No concepts available.

Metadata

Type: preprint
Language: en
Landing Page: http://arxiv.org/abs/2510.10994
PDF: https://arxiv.org/pdf/2510.10994
OA Status: green
OpenAlex ID: https://openalex.org/W4417100892

All OpenAlex metadata

Raw OpenAlex JSON

OpenAlex ID: https://openalex.org/W4417100892

Canonical identifier for this work in OpenAlex
DOI: https://doi.org/10.48550/arxiv.2510.10994

Digital Object Identifier
Title: DeepResearchGuard: Deep Research with Open-Domain Evaluation and Multi-Stage Guardrails for Safety

Work title
Type: preprint

OpenAlex work type
Language: en

Primary language
Publication year: 2025

Year of publication
Publication date: 2025-10-13

Full publication date if available
Authors: Henry Peng Zou, Dongyuan Li, Andrea Zangari, Jing Guo, Chunyan Miao, Liang He, Renhe Jiang, Philip S. Yu

List of authors in order
Landing page: https://arxiv.org/abs/2510.10994

Publisher landing page
PDF URL: https://arxiv.org/pdf/2510.10994

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/2510.10994

Direct OA link when available
Cited by: 0

Total citation count in OpenAlex

Full payload

id	https://openalex.org/W4417100892
doi	https://doi.org/10.48550/arxiv.2510.10994
ids.doi	https://doi.org/10.48550/arxiv.2510.10994
ids.openalex	https://openalex.org/W4417100892
fwci
type	preprint
title	DeepResearchGuard: Deep Research with Open-Domain Evaluation and Multi-Stage Guardrails for Safety
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:2510.10994
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/2510.10994
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/2510.10994
locations[1].id	doi:10.48550/arxiv.2510.10994
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.2510.10994
indexed_in	arxiv, datacite
authorships[0].author.id	https://openalex.org/A5109737659
authorships[0].author.orcid	https://orcid.org/0009-0003-5259-4998
authorships[0].author.display_name	Henry Peng Zou
authorships[0].author_position	middle
authorships[0].raw_author_name	Zou, Henry Peng
authorships[0].is_corresponding	False
authorships[1].author.id	https://openalex.org/A5047165324
authorships[1].author.orcid	https://orcid.org/0000-0002-4462-3563
authorships[1].author.display_name	Dongyuan Li
authorships[1].author_position	middle
authorships[1].raw_author_name	Li, Dongyuan
authorships[1].is_corresponding	False
authorships[2].author.id	https://openalex.org/A5038611842
authorships[2].author.orcid	https://orcid.org/0000-0001-8711-3072
authorships[2].author.display_name	Andrea Zangari
authorships[2].author_position	last
authorships[2].raw_author_name	Zangari, Angelo
authorships[2].is_corresponding	False
authorships[3].author.id	https://openalex.org/A5007013151
authorships[3].author.orcid	https://orcid.org/0000-0002-2595-7314
authorships[3].author.display_name	Jing Guo
authorships[3].author_position	middle
authorships[3].raw_author_name	Guo, Jizhou
authorships[3].is_corresponding	False
authorships[4].author.id	https://openalex.org/A5100382077
authorships[4].author.orcid	https://orcid.org/0000-0002-0300-3448
authorships[4].author.display_name	Chunyan Miao
authorships[4].author_position	middle
authorships[4].raw_author_name	Miao, Chunyu
authorships[4].is_corresponding	False
authorships[5].author.id	https://openalex.org/A5101826239
authorships[5].author.orcid	https://orcid.org/0000-0003-2690-2638
authorships[5].author.display_name	Liang He
authorships[5].author_position	middle
authorships[5].raw_author_name	He, Langzhou
authorships[5].is_corresponding	False
authorships[6].author.id	https://openalex.org/A5040449880
authorships[6].author.orcid	https://orcid.org/0000-0003-2593-4638
authorships[6].author.display_name	Renhe Jiang
authorships[6].author_position	middle
authorships[6].raw_author_name	Jiang, Renhe
authorships[6].is_corresponding	False
authorships[7].author.id	https://openalex.org/A5036357902
authorships[7].author.orcid	https://orcid.org/0000-0002-3491-5968
authorships[7].author.display_name	Philip S. Yu
authorships[7].author_position	middle
authorships[7].raw_author_name	Yu, Philip S.
authorships[7].is_corresponding	False
has_content.pdf	True
has_content.grobid_xml	False
is_paratext	False
open_access.is_oa	True
open_access.oa_url	https://arxiv.org/pdf/2510.10994
open_access.oa_status	green
open_access.any_repository_has_fulltext	False
created_date	2025-10-15T00:00:00
display_name	DeepResearchGuard: Deep Research with Open-Domain Evaluation and Multi-Stage Guardrails for Safety
has_fulltext	False
is_retracted	False
updated_date	2025-12-08T09:52:00.751972
primary_topic
cited_by_count	0
locations_count	2
best_oa_location.id	pmh:oai:arXiv.org:2510.10994
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/2510.10994
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/2510.10994
primary_location.id	pmh:oai:arXiv.org:2510.10994
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/2510.10994
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/2510.10994
publication_date	2025-10-13
publication_year	2025
referenced_works_count	0
abstract_inverted_index.a	87, 122, 129
abstract_inverted_index.In	118
abstract_inverted_index.To	80
abstract_inverted_index.We	100
abstract_inverted_index.an	150
abstract_inverted_index.as	44, 58
abstract_inverted_index.be	222
abstract_inverted_index.by	162, 173
abstract_inverted_index.in	7, 33, 69
abstract_inverted_index.of	48, 54, 96, 121, 156
abstract_inverted_index.or	71
abstract_inverted_index.to	20
abstract_inverted_index.we	84, 126, 193
abstract_inverted_index.6%.	163
abstract_inverted_index.Our	136
abstract_inverted_index.The	164, 219
abstract_inverted_index.and	26, 37, 63, 98, 110, 113, 146, 181, 187, 201
abstract_inverted_index.are	31
abstract_inverted_index.but	50
abstract_inverted_index.can	221
abstract_inverted_index.for	132
abstract_inverted_index.key	115
abstract_inverted_index.may	67
abstract_inverted_index.out	175
abstract_inverted_index.the	77, 119, 168, 179
abstract_inverted_index.via	224
abstract_inverted_index.web	12
abstract_inverted_index.Deep	0
abstract_inverted_index.They	40
abstract_inverted_index.This	65
abstract_inverted_index.code	220
abstract_inverted_index.deep	15, 133
abstract_inverted_index.five	114
abstract_inverted_index.from	11
abstract_inverted_index.have	3
abstract_inverted_index.into	76
abstract_inverted_index.most	169
abstract_inverted_index.plan	180
abstract_inverted_index.rate	109, 154, 161
abstract_inverted_index.show	194
abstract_inverted_index.such	57
abstract_inverted_index.that	195, 204
abstract_inverted_index.with	93
abstract_inverted_index.LLMs,	141
abstract_inverted_index.While	14
abstract_inverted_index.being	74
abstract_inverted_index.block	206
abstract_inverted_index.e.g.,	106
abstract_inverted_index.exact	45
abstract_inverted_index.final	78
abstract_inverted_index.found	223
abstract_inverted_index.guard	166
abstract_inverted_index.input	165
abstract_inverted_index.match	46
abstract_inverted_index.rate,	112
abstract_inverted_index.shown	4
abstract_inverted_index.spans	138
abstract_inverted_index.these	82
abstract_inverted_index.treat	42
abstract_inverted_index.while	158, 178, 210
abstract_inverted_index.18.16%	157
abstract_inverted_index.across	103
abstract_inverted_index.assess	101
abstract_inverted_index.depth,	62
abstract_inverted_index.guards	183
abstract_inverted_index.issues	23
abstract_inverted_index.rates.	218
abstract_inverted_index.report	55, 116, 213
abstract_inverted_index.result	68
abstract_inverted_index.risks,	177
abstract_inverted_index.safety	124
abstract_inverted_index.source	188
abstract_inverted_index.GPT-4o,	143
abstract_inverted_index.Through	190
abstract_inverted_index.absence	120
abstract_inverted_index.address	21, 81
abstract_inverted_index.aspects	53
abstract_inverted_index.average	151
abstract_inverted_index.complex	22
abstract_inverted_index.content	208
abstract_inverted_index.crucial	52
abstract_inverted_index.cycles,	28
abstract_inverted_index.defense	107, 152
abstract_inverted_index.diverse	139
abstract_inverted_index.enables	197
abstract_inverted_index.enhance	184
abstract_inverted_index.harmful	207
abstract_inverted_index.issues,	83
abstract_inverted_index.obvious	176
abstract_inverted_index.quality	56, 214
abstract_inverted_index.report.	79
abstract_inverted_index.reports	10
abstract_inverted_index.safety.	64, 135
abstract_inverted_index.sources	73
abstract_inverted_index.success	108, 153
abstract_inverted_index.through	24
abstract_inverted_index.without	215
abstract_inverted_index.accuracy	47
abstract_inverted_index.achieves	149
abstract_inverted_index.breadth,	61
abstract_inverted_index.citation	185
abstract_inverted_index.defenses	203
abstract_inverted_index.existing	29
abstract_inverted_index.metrics,	105
abstract_inverted_index.multiple	104
abstract_inverted_index.o4-mini.	147
abstract_inverted_index.overlook	51
abstract_inverted_index.planning	25
abstract_inverted_index.provides	167
abstract_inverted_index.reducing	159
abstract_inverted_index.reports.	99
abstract_inverted_index.research	1, 16, 27, 134, 182
abstract_inverted_index.sources.	13
abstract_inverted_index.suitable	123
abstract_inverted_index.benchmark	131
abstract_inverted_index.deficient	32
abstract_inverted_index.excessive	216
abstract_inverted_index.extensive	191
abstract_inverted_index.featuring	90
abstract_inverted_index.filtering	174
abstract_inverted_index.framework	89
abstract_inverted_index.hazardous	70
abstract_inverted_index.improving	212
abstract_inverted_index.including	142
abstract_inverted_index.introduce	85, 127
abstract_inverted_index.malicious	72
abstract_inverted_index.oversight	66
abstract_inverted_index.possesses	17
abstract_inverted_index.potential	19
abstract_inverted_index.promising	5
abstract_inverted_index.typically	41
abstract_inverted_index.benchmark,	125
abstract_inverted_index.coherence,	60
abstract_inverted_index.discipline	186
abstract_inverted_index.evaluation	35, 43, 95, 137, 200
abstract_inverted_index.four-stage	91
abstract_inverted_index.frameworks	2, 30
abstract_inverted_index.integrated	75
abstract_inverted_index.procedures	36
abstract_inverted_index.protection	172
abstract_inverted_index.references	97
abstract_inverted_index.safeguards	92
abstract_inverted_index.stage-wise	130
abstract_inverted_index.sufficient	34
abstract_inverted_index.dimensions.	117
abstract_inverted_index.early-stage	171
abstract_inverted_index.effectively	205
abstract_inverted_index.improvement	155
abstract_inverted_index.open-domain	94, 199
abstract_inverted_index.performance	102
abstract_inverted_index.significant	18
abstract_inverted_index.stage-aware	202
abstract_inverted_index.substantial	170
abstract_inverted_index.DRSAFEBENCH,	128
abstract_inverted_index.DeepSeek-v3,	145
abstract_inverted_index.capabilities	6
abstract_inverted_index.credibility,	59
abstract_inverted_index.credibility.	189
abstract_inverted_index.experiments,	192
abstract_inverted_index.over-refusal	111, 160, 217
abstract_inverted_index.propagation,	209
abstract_inverted_index.protections.	39
abstract_inverted_index.synthesizing	8
abstract_inverted_index.comprehensive	9, 88, 198
abstract_inverted_index.stage-specific	38
abstract_inverted_index.systematically	211
abstract_inverted_index.state-of-the-art	140
abstract_inverted_index.DEEPRESEARCHGUARD	148, 196
abstract_inverted_index.Gemini-2.5-flash,	144
abstract_inverted_index.DEEPRESEARCHGUARD,	86
abstract_inverted_index.question-answering,	49
abstract_inverted_index.https://github.com/Jasonya/DeepResearchGuard.	225
cited_by_percentile_year
countries_distinct_count	0
institutions_distinct_count	8
citation_normalized_percentile