Recent advances in video generation call for reliable detection grounded in explicit forensic evidence. However, current methods primarily depend on single-pass chain-of-thought reasoning that overlooks brief or localized artifacts, or on predefined multi-agent workflows that offer limited flexibility in selecting relevant inspections. Even with adaptive investigation, optimizing the investigation policy alone does not directly improve the forensic capabilities of fixed experts. We introduce WeVisForensics, an agentic framework in which a main investigator agent coordinates a self-evolving library of forensic inspector sub-agents spanning visual and audio cues. The investigator learns which artifact type to examine, over which window, and when to conclude, and these interactions provide shared feedback for both levels of agents. Our dual-timescale policy optimization strategy, DTPO, closes this loop through fast reinforcement learning of the investigator and slower refinement of the inspectors, guided by annotated artifact intervals, to report subtle artifacts they once dismissed. Furthermore, we curate ForenCue-30K, combining role-specific forensic annotations and interaction trajectories to supervise individual inspections and their coordination, with an accompanying reinforcement-learning pool. Experiments show that WeVisForensics averages 95.2% detection accuracy across in-domain, out-of-domain and agentic benchmarks, 9.0% above the strongest prior detector, while grounding its verdicts in localized evidence.
Figures & tables
Figure 2: Overview of WeVisForensics. Top: the investigator πθ repeatedly calls an inspector st on a temporal scope ut , reads its report zt , and issues a verdict with the evidence chain. Middle: DTPO updates the investigator on every batch, while the trainable inspectors self-evolve on a slower timescale. Bottom: accuracy on the four evaluation benchmarks against representative detectors.
Figure 3: WeVisForensics and its training strategy, DTPO. The investigator rolls out with a fixed FSL snapshot and is updated on every batch with outcome rewards and CLUE. Grounded calls ( DOPD ) queue by role in the interaction pool, from which the slow process distills privileged EMA teachers into students, revises skill cards, and refreshes the FSL every H steps.
Figure 4: Overview of ForenCue-30K. Curated videos are annotated with role-specific artifact evidence and converted into sub-agent records and interaction trajectories, and an RL pool supports DTPO. Examples show artifacts targeted by each sub-agent.
Appear.
Phys.
OCR
Comm.
Audio
Eligible
2,328
2,328
2,328
2,328
2,330
∣Gs∣
2,121
1,829
1,146
558
500
Base (%)
91.1
78.6
49.2
24.0
21.5
Table 3: Label counts behind sub-agent invocation precision (Table 5 ). Eligible counts the generated videos scored for each role, ∣Gs∣ the videos that call for role s , and Base the precision of invoking tool s on every eligible video, which is also the precision of the fixed-workflow ARGUS on its two observer roles.
OOD generator categories
Schedule
Hailuo
Jimeng
Kling
Seedance
Sora2
Wan2.5
Mean
DTPO ( H=20 )
88.9
82.9
89.6
98.6
82.4
85.6
88.0
Every step
84.1
82.3
87.0
98.2
80.1
84.3
86.0
Staged
85.5
81.5
84.9
97.8
80.2
83.9
85.6
Δ vs. every step
↑ 4.8
↑ 0.6
↑ 2.6
↑ 0.4
↑ 2.3
↑ 1.3
↑ 2.0
Δ vs. staged
↑ 3.4
↑ 1.4
↑ 4.7
↑ 0.8
↑ 2.2
↑ 1.7
↑ 2.4
Table 7: Update cadence of the forensic library. Left: accuracy (%) on unseen OOD generators comparing refreshing the library every H=20 steps as in DTPO, refreshing it on every rollout batch, and a staged schedule that first evolves the library and then freezes it while training the main agent. All start from the same SFT agent and library and are evaluated at the same step, and Δ is DTPO minus each alternative. Right: training reward against wall-clock time.
Role
FSL
GB
GB++
MintVid
AgenticCue
Appearance
Frozen
97.3
97.1
79.4
95.4
Evolving
98.1 ↑ 0.8
97.5 ↑ 0.4
86.4 ↑ 7.0
98.7 ↑ 3.3
Physical
Frozen
66.5
85.1
72.7
89.4
Evolving
97.8 ↑ 31.3
97.3 ↑ 12.2
91.9 ↑ 19.2
98.5 ↑ 9.1
OCR
Frozen
89.3
76.7
56.6
77.9
Evolving
92.7 ↑ 3.4
88.5 ↑ 11.8
58.9 ↑ 2.3
84.0 ↑ 6.1
Table 8: Sub-agent evidence quality. Sub-agent F1 (%) against role-level annotations. A call counts when its role is annotated and its window overlaps the interval. Both libraries answer the same matched calls. Frozen: SFT library in GRPO. Evolving: refined by DTPO. OCR excludes text-presence routing.
Figure 5: Self-evolution of the forensic library over training. (a) Sub-agent verdicts on training calls, scored against role-level annotations. Left: moving averages of the miss rate on generated videos and of the verdict F1 over LoRA refreshes (minor ticks). Middle: role-level error per sub-agent across the early, middle, and late phases of training. Right: evidence F1 of the frozen library and of the evolved library at the early, middle, and final checkpoints on four test benchmarks, each spanning the appearance, physical, and OCR axes clockwise. (b) Accepted edits, with one Report and one Ignore rule each. The gray card shows candidate rules that failed the acceptance test. (c) Three generated test videos inspected by the three checkpoints, with verbatim sub-agent excerpts and the agent’s final answer.
Figure 6: Qualitative comparison on a generated video with warping handwritten text. Left: a) The CoT-based VideoVeritas [ 36 ] dismisses the garbled text as stylistic without inspecting it. b) In the fixed-workflow ARGUS [ 56 ] , texture, lighting, and motion reports outweigh the detected text deformation. Right: The SFT or DTPO-trained main agent paired with the SFT or evolved sub-agents. The evolved sub-agents flag the glyph anomalies, and only the trained main agent, which moves its OCR inspection to the warped text at 7 s, turns them into a fake verdict.
Investigator \ Inspectors
SFT
DTPO
SFT
91.3
92.0
DTPO
93.2
95.2
Table 9: Contribution of each level to investigator–inspector co-evolution. The investigator and inspectors before (SFT) or after (DTPO) co-evolution are paired in all four combinations, reporting average accuracy (%) over GB, GB++, MintVid, and AgenticCue. Evolving either level alone helps, and evolving both gives the largest gain.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 10: Quantitative human verification audit of ForenCue-30K. (a) Sub-agent role annotations over 100 generated videos ( 340 video–role annotations), with numbers above bars giving the at-least-partially-supported rate. (b) Main-agent interaction trajectories over 100 videos ( 50 generated and 50 real, 209 tool calls), with numbers above bars giving trajectory-level tool-selection appropriateness (left of the dotted line) and call-level precision/recall (right).
Preferred
Missed
Flawed
Sub-agent
Pairs
S
E
S → E
S → E
Appearance
58
6
21
32.8 → 1.7
34.5 → 8.6
Physical
62
6
21
27.4 → 1.6
29.0 → 6.5
OCR
63
7
23
33.9 → 4.8
36.5 → 11.1
Commonsense
17
1
0
0.0 → 0.0
0.0 → 0.0
All
200
20
65
28.6 → 2.5
30.5 → 8.0
Appendix
Table 10: Paired human review of sub-agent evidence. Each pair answers the same inspection with the SFT (S) and evolved (E) sub-agent. Preferred counts pairs where one response gives more accurate evidence, omitting ties and pairs where both are unreliable. Missed: a clearly visible relevant anomaly is not reported. Flawed: the response misses a visible anomaly or reports one that is unsupported, out of role, or mislocalized. Rates in %.
Figure 11: Human review interfaces. (a) The reviewer filters the queue by source and review status, watches the full clip, and records whether an audio inconsistency is present, absent, or inconclusive, along with the supporting time interval. (b) The reviewer inspects the stated window and compares two anonymized sub-agent responses to the same inspection, marking the more accurate evidence and flagging unsupported claims and missed anomalies in each.
Figure 12: Comparison of AI-generated video detection paradigms. a) CoT-based reasoning uses textual rationales. b) Multi-agent workflows follow fixed execution structures. c) WeVisForensics adaptively coordinates sub-agents and evolves them together with the main policy.
Figure 13: Qualitative comparison of investigation paradigms. All three approaches classify the generated video as fake. The CoT-based detector makes unsupported motion claims, and the fixed-workflow detector aggregates predefined observer reports. The adaptive agent requests localized appearance and OCR inspections and grounds its verdict in changing emblem contours and degraded apron lettering.
Figure 14: Complete evolution comparison. On an AI-generated video, the CoT-based VideoVeritas and the fixed-workflow ARGUS both answer real. Before DTPO (top right), the appearance sub-agent reports no issue and WeVisForensics answers real. After DTPO (bottom right), the same first inspection localizes soft hand and finger boundaries during the opening gesture, and the main agent answers fake.
The rapid evolution of video generation technologies poses a significant challenge to media forensics, as conventional detection methods often fail to generalize beyond their training distributions. To address this, we propose DVAR (Debate-based Video Authenticity Reasoning), a training-free framework that reformulates video detection as a structured multi-agent forensic reasoning process. Moving beyond the paradigm of pattern matching, DVAR orchestrates a competition between a Generative Hypothesis Agent and a Natural Mechanism Agent. Through iterative rounds of cross-examination, these agents defend their respective explanations against abnormal evidence, driving a logical convergence where the truth emerges from rigorous stress-testing. To adjudicate these conflicting claims, we apply Occam's Razor through the Minimum Description Length (MDL) framework, defining an Explanatory Cost to quantify the "logical burden" of each reasoning path. Furthermore, we integrate GenVideoKB, a dynamic knowledge repository that provides high-level reasoning heuristics on generative boundaries and failure modes. Extensive experiments demonstrate that DVAR achieves competitive performance against supervised state-of-the-art methods while exhibiting superior generalization to unseen generative architectures. By transforming detection into a transparent debate, DVAR provides explicit, interpretable reasoning traces for robust video authenticity assessment.
Hongyuan Qi, Feifei Shao, Ming Li +2
Zhejiang University Hangzhou, Zhejiang, China · Guangming Lab Shenzhen, Guangdong, China
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.
AI-generated image detection is increasingly open-world: new image generators produce highly realistic images that make visual artifacts harder to identify. Existing detectors usually rely on a fixed set of forensic cues, so a detector that works well for one generator family may fail on another. We introduce FACT (Forensic Agent with Compiled Tool-use Trajectories), which learns an image-conditioned tool-use policy for forensic analysis. Instead of applying a fixed detector, FACT decides which forensic tools to call, interprets the returned evidence, and stops when sufficient evidence has been collected. FACT follows an Evolve--Distill--Refine pipeline: it evolves an execution-verified forensic skill, compiles the skill into action--observation tool-use trajectories, distills them into a compact agent, and refines the policy with cost-aware GRPO. Across two internal and four public benchmarks, FACT achieves the best performance among all compared methods, including on recent unseen generators, deepfakes, and manipulated images.
Jiaoyang Chen, Bin Hu, Jingyu Hu +3
Lingnan University, Hong Kong · Harbin Institute of Technology, Shenzhen · The University of Hong Kong +2