Verus-Insight: Towards Explainable Artifact Detection in AI-Generated Videos

• Reliable AI-generated video detection matters for content authenticity verification, generation quality evaluation, and provenance and safety analysis.Source: §1 · p. 1
• Existing detectors may overlook subtle spatiotemporal abnormalities, mistake normal phenomena for artifacts, or reach correct source decisions with inaccurate or incomplete evidence.These weaknesses limit their usefulness for diagnosing specific problems in generated videos.Source: §1 · p. 1
• Verus-Insight structures artifact detection into Inspection Planning, Evidence Verification, and Evidence Aggregation, trained with human-anchored full-process supervision and progressive capability training.Verus-Bench provides fine-grained human annotations of perceptible artifacts for evaluation beyond video-level source prediction.Source: §1 · p. 2
• Experiments show strong performance in fine-grained artifact analysis and video-level detection, with more reliable conversion of relevant inspections into supported findings without more spurious artifact claims.The analyses describe the three capabilities as complementary yet distinct.Source: §1 · pp. 2–3

2.1. Multimodal Large Language Models for Video Understanding and Authenticity Assessment

• General-purpose MLLMs combine visual perception and language reasoning, motivating authenticity assessment; LOKI evaluates synthetic-content detection and anomaly explanation [26], while LAVID adds explicit knowledge [14].Source: §2.1 · p. 3

2.2. Specialized AI-Generated Video Detection and Artifact Analysis.

• Specialized detectors add artifact explanations, but classification and aggregate explanation scores may not verify each claim's semantic accuracy or spatiotemporal grounding, [4],.The cited approaches include classification rewards, human-annotated artifacts, and grounding supervision.Source: §2.2 · p. 3
• Other methods use grounding, similarity, or tool-mediated evidence collection, yet their evaluation does not directly establish whether individual claims comprehensively capture correctly localized human-perceived artifacts [18], [11], [24].Source: §2.2 · p. 3

3.1.1. What Capabilities Make an Effective Artifact Detector?

• Artifact detection can fail through missed inspection, misinterpreted evidence, or correct source prediction without recovering salient artifacts; the stages are Inspection Planning, Evidence Verification, and Evidence Aggregation.The stages identify relevant content, verify genuine artifacts with spatial and temporal support when available, and integrate verified evidence into the source prediction.Source: §3.1.1 · pp. 3–4

3.1.2. How to Strengthen Intermediate Capabilities?

• Sample-specific, outcome-neutral rubrics support Inspection Planning, and combining rubric warm-up with explicit rubric generation performs best across all three reported metrics.The authors preserve the generated rubric as the interface between planning and verification.Source: Table 1 · p. 5, §3.1.2 · p. 5
• Region tracking combined with explicit spatiotemporal grounding achieves the best overall reported performance for Evidence Verification.The authors adopt region tracking for perceptual warm-up and retain explicit grounding during verification.Source: Table 1 · p. 5, §3.1.2 · p. 5
• The design studies recommend targeted supervision and explicit representations within the full detection process to provide evidence for subsequent aggregation.Source: §3.1.2 · p. 5

3.2. Rubric-Guided Evidence Verification: Supervision Construction

• Human artifact annotations are transformed into sample-specific, outcome-neutral criteria linked to verification targets that retain the observed abnormality and available spatial and temporal evidence.Source: §3.2 · p. 6
• Verified-normal criteria are added to prevent criteria from revealing artifact status, and criterion-verification pairs are shuffled while preserving correspondence in the full-process trajectory.The trajectory also includes a label-neutral description, inspection domains, and an aggregation step that introduces no additional artifact claims.Source: §3.2 · pp. 6–7

3.3. Progressive Three-Stage Training

• Progressive training moves from joint tracking and rubric-planning warm-up to full-process SFT, then GRPO optimization of the complete detection process.Full-process SFT supervises descriptions, criteria, verification results, aggregation, and the final verdict; reinforcement learning jointly encourages correct source prediction and faithful artifact recovery.Source: §3.3 · p. 7
• The reinforcement-learning reward is gated by valid output and correct source verdict, then rewards artifact recovery in generated videos or avoidance of spurious predictions in real videos.F1find measures matching between predicted artifacts and reference annotations, while Npred is the number of predicted artifacts.Source: §3.3 · p. 7

• Verus-Bench targets fine-grained artifact-detection evaluation beyond video-level source prediction.Source: §4 · p. 7
• Its real videos come from NExT-QA [25], OneThinker-STVG [3], DiDeMo [1], and an Internet collection; generated videos are independently collected and sample-disjoint from training and development data.Source: §4 · p. 7
• Dense frame-by-frame annotations capture salient human-perceived artifacts with concise observations, temporal extents, and frame-wise bounding boxes, followed by manual quality review.Source: §4 · p. 7

5.1. Experimental Setup

• Video-level detection uses Artifact-Bench, RewardData, ViF-Bench, and Verus-Bench; fine-grained analysis uses Verus-Bench and RewardData, with ViF-CoT-4K subset for generalization.The baselines include proprietary, open-source general-purpose, and open-source specialized models, evaluated with a common scoring pipeline where applicable.Source: §5.1 · p. 8

5.2. Main Results

• 48.81 to 94.79 BA on RewardData and 66.67 to 89.06 BA on Verus-Bench; versus Qwen3-VL-8B, Artifact-Bench T1 rises from 50.80 to 66.60.Among open-source models, Verus-Insight achieves the best results on RewardData, Verus-Bench, and Artifact-Bench T1 and T3, while remaining competitive on other benchmarks.Source: Table 2 · p. 8, §5.2 · p. 9
• On RewardData, Finding F1 rises from 3.38 to 35.42 and Plan F1 from 18.67 to 31.33 versus Qwen3-VL-8B; Verus-Insight outperforms evaluated specialized detectors on both benchmarks.On Verus-Bench, Plan F1 increases from 17.08 to 24.70 and Finding F1 from 7.56 to 15.97. On RewardData, Finding precision and recall rise from 1.87/18.13 to 33.66/37.36.Source: Table 3 · p. 9, §5.2 · p. 9

5.3.1. Ablation Study

• The full configuration achieves the highest Plan F1 and Finding F1 on both benchmarks; removing the Stage-II rubric lowers both despite slightly improving BA.Removing the Stage-II rubric changes Plan/Finding F1 from 24.70/15.97 to 21.54/13.76 on Verus-Bench and from 31.33/35.42 to 26.57/29.35 on RewardData. Removing rubric warm-up degrades all three metrics, and removing perceptual warm-up or reinforcement learning consistently reduces BA, Plan F1, and Finding F1.Source: Table 4 · p. 10, §5.3.1 · p. 10

Q1: Inspection Planning, Evidence Verification, and Evidence Aggregation Are Distinct Capabili-

• On RewardData, comparable Plan F1 for Verus-Insight and Gemini 3.6 Flash (31.33 vs. 33.13) accompanies higher Finding F1 (35.42 vs. 25.54); conversion rises to 75.58% from 30.48%.Verus-Insight also raises conversion from 45.16% to 54.88% on Verus-Bench despite lower Plan recall than its backbone; it has the lowest strict artifact false-positive rate on real RewardData videos.Source: Q1: Inspection Planning, Evidence… · pp. 10–11, Figure 4 · p. 11, Figure 5 · p. 11
• On ViF-CoT-4K subset, Verus-Insight reaches 35.27 Plan F1 and 14.02 Finding F1; on Verus-Bench it leads compared specialized detectors in Any-Hit and coverage with at least three artifacts and recall across extent bins.RewardData shows overall coverage gains, though improvements are not uniform across all extent bins.Source: Q1: Inspection Planning, Evidence… · pp. 11–12, Figure 6 · p. 12

6. Conclusion

• Verus-Insight more reliably converts relevant inspections into accurate artifact findings while limiting spurious claims, and treats inspection coverage, artifact recovery, and source prediction as related but distinct.The authors motivate future artifact detectors as evidence-grounded inspection systems linking content-specific inspection, concrete visual evidence, and final source decisions.Source: §6 · p. 12
• Human-anchored full-process supervision and progressive capability training improve both fine-grained artifact recovery and video-level detection.The framework decomposes detection into Inspection Planning, Evidence Verification, and Evidence Aggregation.Source: §6 · p. 12
• Verus-Bench provides fine-grained human annotations of perceptible artifacts and their spatiotemporal evidence for evaluating artifact detection.Source: §6 · p. 12
• The design studies highlight combining targeted capability supervision with explicit inspection criteria and spatiotemporal grounding.Source: §6 · p. 12

A.1.1. Verus-Annotation and Training Data

• Human annotators record artifact categories, abnormality descriptions, temporal intervals, and frame-wise bounding boxes in Verus-Annotation.Its training and development splits are provided, with only the training split used to construct training supervision.Source: §A.1.1 · p. 15
• Training instances combine Verus-Annotation training data with RewardData videos and annotations, converting artifacts into outcome-neutral inspection criteria linked to abnormal observations and spatiotemporal evidence.Verified-normal checks supplement artifact-anchored criteria and provide inspection targets for real videos; supervision includes a video description, criteria, verification results, and an aggregated source verdict.Source: §A.1.1 · p. 15
• 8,411 enriched instances comprise 6,411 from Verus-Annotation and 2,000 from RewardData’s training split.Table 5 summarizes source composition and supervision counts after enrichment.Source: §A.1.1 · p. 15, Table 5 · p. 15
• Figure 7 shows 1–9 Rubric criteria and 0–7 artifact checks per enriched instance; each Rubric criterion links to a verification record.The figure presents distributions of Rubric criteria, artifact checks, and normal checks across 8,411 instances.Source: §A.1.1 · p. 15, Figure 7 · p. 16

A.1.2. Verus-Bench

• Generated videos are independently collected from the Internet, sample-disjoint from training and development, and manually quality-reviewed; real videos come from the sources described in Section 3.2.The benchmark annotations use the same protocol as Verus-Annotation.Source: §A.1.2 · p. 16
• Verus-Bench contains 192 videos, evenly split between real and AI-generated, with 326 manually annotated generated-video findings across 18 artifact categories.Each generated video has 1–5 findings, each with an observation, temporal extent, and frame-wise bounding boxes.Source: §A.1.2 · p. 16, Figure 9 · p. 17
• Sampled inputs contain 20–60 frames, with a median of 20, spanning 4.75–14.75 seconds, with a median span of 4.75 seconds.Figures 8 and 9 summarize video-level statistics and the artifact-category distribution.Source: §A.1.2 · p. 16, Figure 8 · p. 16, Figure 9 · p. 17

A.1.3. Development data

• Analysis-Dev contains 576 videos from Verus-Annotation development, RewardData validation, and the ViF-CoT source training pool, with 192 videos per source and 288 real and 288 AI-generated videos overall.The selected ViF-CoT videos are held out for development evaluation and separate from the ViF-Bench test set.Source: §A.1.3 · p. 17
• Analysis-Dev is sample-disjoint from this study’s training data and final test sets, checked using parent-video identities, paths, and frame content.Source: §A.1.3 · p. 17
• The set reuses existing annotations, totaling 545 ground-truth findings across the three subsets, and evaluates Plan F1, Finding F1, and video-level Balanced Accuracy.AI-generated RewardData videos require original artifact explanations, while selected ViF-CoT videos each require exactly one explicitly annotated finding; annotations are not regenerated or manually split.Source: §A.1.3 · p. 17

A.2.1. Definitions of O, L, T, R, and G

• O describes and categorizes visible abnormalities; L localizes described artifacts; T predicts target boxes at query times; R plans content-specific inspection criteria; G grounds observations in intervals and boxes or tracks.Source: §A.2.1 · pp. 17–18

A.2.2. Training Configuration

• Both SFT stages update the language model, visual encoder, and multimodal merger; RL uses a frozen Stage-II KL reference, balanced real and generated video sampling, and development-split checkpoint selection.SFT uses Adam with β1 = 0.9, β2 = 0.95, ε = 10^-8, weight decay 0.1, gradient clipping 1.0, and BF16; reward evaluation uses λ = 0.2 and a Gemini 3.1 Pro judge.Source: §A.2.2 · pp. 18–19

A.3.1. Overall Artifact Coverage

• Verus-Insight achieves the highest scores among evaluated open-source models on both benchmarks, raising Any-Hit from 33.68% to 46.32% on Verus-Bench and from 19.21% to 45.03% on RewardData versus Qwen3-VL-8B.Any-Hit is the percentage of videos with at least one recovered annotated artifact.Source: Table 8 · p. 19, §A.3.1 · p. 19

A.3.2. Artifact Recovery on RewardData

• 68.00% Any-Hit on multi-artifact RewardData videos exceeds VideoVeritas at 44.00% and Skyra at 40.00%, while macro coverage is 31.33% versus VideoVeritas at 30.67%.Verus-Insight matches VideoVeritas on short and medium artifacts, does better on long artifacts and medium and large regions, but trails it on small regions.Source: §A.3.2 · pp. 19–20, Figure 10 · p. 20

A.3.3. Qualitative Failure Analysis

• In the examples, planning misses a plank’s spontaneous movement, while verification inspects finger-clay interaction but judges the annotated finger-clay merging natural.The verification example is a reference-relative interpretation error, not an absence of inspection, and is not independent human adjudication.Source: Model response (verbatim excerpts) · p. 21, §A.3.3 · p. 21, Reference annotation (verbatim) · pp. 21–22
• In the aggregation example, the model identifies distorted text as an artifact but attributes it to post-production and predicts real, showing that a source-verdict error can coexist with artifact recovery.The response yields eight Plan checks and one affirmative Finding, which can match only one reference under one-to-one assignment.Source: Reference annotation (verbatim) · pp. 22–23, §A.4.3 · p. 25

A.4.2. Semantic matching and micro-averaged metrics

• Evaluation separates Plan checks from affirmative Findings, matches each independently by target and specific property or phenomenon, and uses maximum one-to-one assignment to compute counts.Normal judgments, negations, questions, suspicions, and uncertain alternatives are excluded as Findings; unmatched affirmative Findings count as false positives, including on reference-empty videos.Source: Two evaluation objects from one… · p. 23, §A.4.2 · p. 24

A.6. Training Cases: Plan, Check, and Verdict

• Training targets link label-neutral rubric criteria to verification outcomes and aggregate them into verdicts, while rubric construction preserves finding identity and temporal scope.Normal-check candidates are proposed without human findings or source labels and retained only after independent visual verification; T, L, and O share a prompt with task-specific user prompts and normalized [0, 1000] box coordinates.Source: §A.5.1 · p. 25, P1 | Rubric generation · p. 25, §A.5.4 · p. 30

1. Plan | Inspection Planning

• In the generated-sign case, distorted “Calex” lettering supports a fake verdict, while text and scratches remain stable and the scratch rubric checks persistence across frames.The rubric calls for matching distinctive scratches across early, middle, and late frames and accounting for visibility changes from illumination, focus, and camera angle.Source: §1 · pp. 32–33

r2 | Rubric criterion

• For the baby-and-toy real-video case, checks target toy integrity, head support, hand-loop grip, and gaze continuity; all four linked observations are normal and the verdict is real.The criteria check dangling-part consistency, the baby’s head-to-thigh support, grip continuity, and whether the baby’s gaze tracks the toy smoothly.Source: §1 · p. 33, r2 | Rubric criterion · p. 34