Video large multimodal models have improved at identifying cross-frame visual evidence and conducting open-web retrieval, as reflected in recent video deep research (VDR) benchmarks. However, existing VDR benchmarks primarily evaluate tool-use trajectories and final responses without verifying whether a model actually uses the temporal video information or the retrieved web evidence. To fill this gap, we introduce \method{}, an evidence-grounded benchmark with three features. First, it combines human review with restricted-access probing across multiple models to exclude questions answerable from parametric knowledge or a single information source. Second, each question carries structured video-to-web evidence annotations that link the relevant video segments to their supporting web evidence. Third, its evaluation protocol separately scores video grounding, web evidence grounding, and final-answer correctness. Across eight multimodal agent systems, the best reaches 76.0% answer accuracy, yet none recovers the complete required web evidence for more than 8% of questions. Our benchmark is publicly available at https://huggingface.co/datasets/hh721/vdr_temp.
Figures & tables
Figure 1: Failure cases in video deep research (VDR). Case 1 bypasses the video, and Case 2 bypasses web retrieval, illustrating variation in human and model judgments of modality dependence. Case 3 reduces the task to single-frame entity recognition rather than video-guided search. Case 4 produces the correct final answer from an incorrect video segment and irrelevant web evidence. Green text marks correct answers and correctly identified entities, while red text indicates information leakage.
Basic Construction
Necessity Tests
Evidence Annotations
Benchmark
Seed Direction
Anno.
V/W Tests
Multi-model
Multi-hop
Multi-state
V → W Links
Claim Evid.
HN
Image Deep-Research Benchmarks
BrowseComp-VL ( Geng et al., 2025 )
Text → Image
Hybrid
✗
✗
✓
N/A
∘
✗
✗
MMSearch ⋆ ( Jiang et al., 2024 )
Query-first
Manual
✗
✓
∘
N/A
∘
∘
✗
VDR-Bench ( Zeng et al., 2026 )
Image-first
Hybrid
∘
∘
✓
N/A
∘
∘
✗
Video Deep-Research Benchmarks
Table 1: Comparison of benchmark construction, necessity tests, and evidence annotations. Seed Direction gives the order in which the task is constructed, and Annotation (Anno.) identifies manual or hybrid construction. Video-only/web-only tests (V/W Tests) check whether either modality alone is sufficient. Multi-model indicates testing across multiple models without search access. Multi-hop and Multi-state indicate multiple reasoning steps and temporally distinct video states, respectively. Video-to-web links (V → W Links) connect localized video evidence to web retrieval conditions. Claim-level evidence (Claim Evid.) records supporting text from web sources, and hard negatives (HN) are evidence-backed incorrect candidates. ✓ , ∘ , and ✗ denote full, partial, and absent support; N/A denotes not applicable.
Figure 2: Four-stage construction procedure for GroundedVDR . The stages are (1) video selection, (2) video evidence mining, (3) video-to-web task curation, and (4) verification and release. Questions are retained only when the annotated video evidence constrains the subsequent web search and the supporting web claims pass automated checks and human review.
Figure 3: Evidence structure of GroundedVDR . (a) Joint distribution of the numbers of required video states and web claims; each cell reports the number of questions. (b) Distribution of temporally localized video evidence units required per question.
Figure 4: Video capabilities required by GroundedVDR . (a) Supporting capabilities associated with each primary video deep research (VDR) capability. (b) Number of supporting video capabilities jointly required by each question.
System
Field Acc.
Answer
Video State
Web Claim
Claude Series
Claude Opus 4.8
61.3
55.0
43.5
16.9
Claude Sonnet 5
61.6
55.0
31.2
17.9
GPT Series
GPT-5.6 Sol
82.0
76.0
37.6
32.0
GPT-5.6 Terra
55.5
47.0
20.5
10.6
Table 2: Main results. Answer metrics evaluate the final response, whereas video and web metrics measure recovery of the required semantic evidence. Field Acc. is the mean per-question proportion of correctly resolved required answer fields, while Answer is the proportion of questions whose required answer fields are all correct. Video State and Web Claim are the mean per-question recalls of required video states and web claims, respectively. Best agent results in each column are shown in bold; the human-expert baseline is reported separately for reference.
Video Evidence
Web Evidence
System
Complete
Partial
None
Complete
Partial
None
GPT-5.6 Sol
21
36
43
8
53
39
Claude Opus 4.8
24
39
37
4
34
62
Claude Sonnet 5
15
34
51
1
42
57
GPT-5.6 Terra
7
28
65
1
27
72
Grok 4.6
8
25
67
3
28
69
Table 3: Partial vs. complete evidence grounding. For each question, video states and web claims are classified as Complete when all required states or claims are supported, Partial when at least one but not all required states or claims are supported, and None when no required state or claim is supported. All values are percentages of the 100 benchmark questions.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Diagnostic
Result
Candidate videos
64
Provisional question blueprints
77
Videos yielding no viable item
23/64 (35.9%)
Earlier questions retrospectively audited
29
Passed temporal necessity
23/29
Passed retrieval necessity
22/29
Appendix
Table 4: Development-stage construction diagnostics. The blueprint pool and retrospective audit precede the frozen benchmark release and are reported only to characterize the construction process.
Models
Agent Platform
Reasoning Setting
Claude Opus 4.8; Claude Sonnet 5
Claude Code
High
GPT-5.6 Sol; GPT-5.6 Terra
Codex
High
Composer 2.5; Grok 4.6
Cursor
High
Qwen3.8-Flash; Qwen3.8-Max
Qoder
Medium
Appendix
Table 5: Models, agent platforms, and reasoning settings used in the evaluation.