Recent advances in video generation have made synthetic videos increasingly realistic, placing greater demands on reliable AI-generated video detection. Existing artifact-based detectors typically identify generation artifacts as evidence for source prediction, yet they may either fail to recognize subtle artifacts or reach the correct source decision without faithfully identifying the evidence underlying it. In this work, we systematically investigate how to design and train effective artifact detectors and formulate artifact detection as a structured process comprising three complementary capabilities: Inspection Planning, which determines what should be examined; Evidence Verification, which verifies whether the inspected content exhibits genuine generation artifacts; and Evidence Aggregation, which integrates the verified evidence into the final source decision. Building on this formulation, we introduce Verus-Insight, trained with human-anchored full-process supervision and progressive capability training to strengthen both fine-grained artifact detection and video-level detection. We further introduce Verus-Bench, a benchmark with fine-grained human annotations of perceptible artifacts for evaluating artifact detection. Extensive experiments across diverse benchmarks demonstrate that Verus-Insight achieves strong performance in both fine-grained artifact analysis and AI-generated video detection, with substantial gains in recovering human-perceived artifacts. Further analyses reveal that the three capabilities are complementary yet distinct, and that Verus-Insight more reliably converts relevant inspection targets into supported artifact evidence without increasing spurious artifact claims.
Figures & tables
Figure 1: Towards explainable and reliable AI-generated video detection. Conventional artifact detectors may make incorrect source predictions or reach the correct conclusion with inaccurate artifact evidence. Verus-Insight instead performs Inspection Planning , Evidence Verification , and Evidence Aggregation , grounding source decisions in explicitly verified, human-perceivable artifact evidence for more explainable and reliable detection.
Figure 2: Distinct behaviors in Artifact Detection. Skyra misses the primary temporal artifact without inspecting the relevant change; VideoVeritas observes the change but misinterprets it as normal styling; BusterX++ reaches the correct source prediction without recovering the salient artifact evidence; Verus-Insight inspects the relevant regions and correctly recovers the salient artifacts.
Warm-up supervision
Explicit interface
Plan F1 ↑
Finding F1 ↑
BA ↑
(a) Inspection Planning
Rubric warm-up
Explicit rubric
No
No
19.76
18.37
78.30
No
Yes
22.63
17.10
77.43
Yes
No
18.75
15.94
84.55
Yes
Yes
23.61
18.68
86.11
Table 1: Development studies for Inspection Planning and Evidence Verification.
Figure 3: Data construction and training pipeline of Verus-Insight. Human artifact annotations and verified-normal checks are organized into criterion-linked verification trajectories for full-process supervision. Training proceeds through tracking and rubric-planning warm-up, full-process SFT, and reinforcement learning.
Model
Artifact-Bench
RewardData
ViF-Bench
Verus-Bench
T1
T2
T3
Proprietary Models
Gemini-3.1-Pro
69.60
57.20
10.29
51.60
52.98
51.04
Gemini-3.6-Flash
66.80
62.00
12.00
90.52
72.22
88.02
GPT-5.5
53.60
52.80
11.43
93.81
62.66
89.06
Open-Source General-Purpose Models
Table 2: Video-level detection results . Artifact-Bench reports BA, accuracy, and exact-match accuracy for T1–T3, respectively; all other benchmarks report BA. Among open-source models, best results are bolded and highlighted, and second best results are underlined.
Model
Verus-Bench
RewardData
Plan
Finding
Plan
Finding
P
R
F1
P
R
F1
P
R
F1
P
R
F1
Proprietary Models
Gemini-3.1-Pro
26.35
27.41
26.87
13.91
26.17
18.16
22.83
46.15
30.55
6.29
42.86
10.97
Gemini-3.6-Flash
31.15
37.07
33.85
25.67
29.91
27.63
22.98
59.34
33.13
16.64
54.95
25.54
GPT-5.5
23.93
45.17
31.28
18.47
32.40
23.53
14.69
65.93
24.02
10.84
54.40
18.08
Table 3: Fine-grained artifact planning and finding results . Plan is evaluated on artifact-positive videos, while Finding is evaluated on the full set.
Verus-Bench
RewardData
Variant
BA
Plan F1
Finding F1
BA
Plan F1
Finding F1
Qwen3-VL-8B Base
66.67
17.08
7.56
48.81
18.67
3.38
Verus-Insight
89.06
24.70
15.97
94.79
31.33
35.42
w/o Stage-II rubric
90.63
21.54
13.76
96.07
26.57
29.35
w/o Perception Warm-up
88.02
22.49
14.40
94.03
28.47
33.16
w/o rubric Warm-up
87.50
22.03
14.29
93.70
27.87
32.24
Table 4: Ablation study of Verus-Insight.
Figure 4: Relationships between stage-wise artifact-detection capabilities. Panels (a,c) compare Plan and Finding F1, while (b,d) compare Finding F1 with video-level BA. Each point denotes a model; dashed diagonals in (a,c) indicate equal Plan and Finding F1.
Figure 5: Diagnostic analysis of artifact recovery. (a) Plan-to-evidence conversion, P(F=1∣P=1) . (b) Strict Artifact FPR on real RewardData videos. (c) Plan and Finding F1 on ViF-CoT-4K subset; Skyra is excluded due to training-data overlap.
Figure 6: Artifact-recovery gains across different conditions. (a,b) Any-Hit rate and artifact coverage across artifact counts; (c,d) artifact recall across temporal and spatial extents. All results use Verus-Bench; n denotes videos in (a,b) and reference artifacts in (c,d).
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Source
Instances
Real / AI-generated
Rubric criteria
Artifact checks
Normal checks
Other checks
Verus-Annotation
6,411
2,875 / 3,536
32,220
10,439
21,577
204
RewardData
2,000
1,000 / 1,000
4,089
1,306
2,782
1
Total
8,411
3,875 / 4,536
36,309
11,745
24,359
205
Appendix
Table 5: Training data composition and supervision counts after enrichment. Each training instance is counted once.
Figure 7: Distributions of Rubric criteria, artifact checks, and normal checks per enriched training instance ( n=8,411 ).
Figure 8: Verus-Bench distributions over 192 videos: (a) reference artifacts per video; (b) sampled input frames.
Figure 9: Artifact-category distribution of Verus-Bench, covering 326 reference findings across 18 observed categories, ordered by frequency.
Source
Source pool
Videos
Real
AI-generated
Artifact findings
VERUS-Annotation
Dev split
192
96
96
323
RewardData
Official val split
192
96
96
126
ViF-CoT
Source training pool
192
96
96
96
Total
576
288
288
545
Appendix
Table 6: Analysis-Dev composition.
Setting
Stage I
Stage II
Stage III
Parameter updates
Full parameters
Full parameters
LoRA ( r=α=32 )
Learning rate
10−5
10−5
10−6
Global batch size
96
96
64 prompts
Training budget
72 updates
176 updates
At most 128 rollout rounds
Learning-rate schedule
5% warm-up, then cosine
5% warm-up, then cosine
Short warm-up, then constant
Appendix
Table 7: Training hyperparameters. SFT batch sizes count sequences, while the RL batch size counts video prompts.
Model
Verus-Bench
RewardData
Any-Hit
Macro
Micro
Any-Hit
Macro
Micro
Proprietary Models
Gemini-3.1-Pro
65.26
26.51
26.17
49.67
46.03
42.86
Gemini-3.6-Flash
70.53
30.98
29.91
62.91
57.51
54.95
GPT-5.5
65.26
32.88
32.40
58.94
55.74
54.40
Open-Source General-Purpose Models
Appendix
Table 8: Additional aggregate artifact-coverage results (%). The common positive cohorts contain 95 Verus-Bench videos and 151 RewardData videos. Macro and Micro denote macro- and micro-averaged coverage. Proprietary models are shown in gray for reference. The best and second-best results among open-source models are highlighted in green and underlined, respectively.
Figure 10: Artifact-recovery analysis on RewardData. (a,b) Any-Hit rate and macro coverage for videos with one or multiple reference artifacts. (c,d) Artifact recall across temporal and spatial extents. n denotes the number of videos in (a,b) and reference artifacts in (c,d). Grouping details are provided in Appendix A.4 .
Panel
Extent
q1
q2
Low / medium / high
Verus-Bench
Temporal
0.100000
0.175000
133 / 85 / 103
Spatial
0.035771
0.160534
107 / 107 / 107
RewardData
Temporal
0.525000
0.983333
60 / 60 / 58
Spatial
0.156621
0.386096
61 / 60 / 61
Appendix
Table 9: Extent-bin thresholds (rounded for display) and reference counts. The hits are semantic recovery events, not localization successes.
School of Cyber Science and Engineering, Southeast University, Nanjing 210096, China · Purple Mountain Laboratories, Nanjing 210000, China · Engineering Research Center of Blockchain Application, Supervision And Management (Southeast University), Ministry of Education, China +2