acceptodds
Under review as a conference paper at ICLR 2027

Neuro-Symbolic Forensic Reasoning for Open-Generator AI-Generated Video Detection

Abstract

AI-generated video detection is critical for information integrity, yet remains hard in open-generator deployment: new generators shift artifact distributions, forensic cues are sparse and heterogeneous, and benchmark shortcuts can overstate progress. We argue that the root cause is the prevailing binary classification formulation f(V ) → 0, 1, which learns generator-specific statistical shortcuts rather than transferable forensic principles, making detectors brittle to unseen generators and costly to retrain. We recast the problem as forensic reasoning: instead of a binary label, the system must produce a verdict together with a grounded evidence chain, f(V ) → (ˆy, E). To instantiate this formulation, we propose FOVEA, a neurosymbolic framework for Forensic Open-generator Video Evidence Acquisition that trains no detector weights. A VLM planner forms forensic hypotheses and routes videos to specialized analysts, whose judgments are grounded by deterministic CV measurements including optical flow, frequency analysis, error-level analysis, temporal tracking, and geometric consistency. A reliability-aware judge fuses heterogeneous evidence via strong-evidence override, confidence-filtered voting, and disagreement-penalized aggregation to produce an interpretable verdict. On GenVideo, across ten unseen generators, FOVEA with Claude Opus 4.6 achieves 97.39% accuracy, outperforming the strongest supervised baseline, NSG-VD, by 5.93 points and reaching 0.99 AUROC, using only a one-time 100-video threshold calibration. The advantage persists after removing visible watermarks, on clips of at least 4 s, and on the GenVidBench benchmark.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer vision, Decision making, reinforcement learning, and control (e.g., hierarchical RL, multi-agent systems, agentic AI)

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jcsSOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents FOVEA (Forensic Open-generator Video Evidence Acquisition), a training-free neuro-symbolic framework for detecting fully AI-generated videos from unseen generators. It combines VLM-based hypothesis formation and adaptive routing with deterministic CV measurements, including optical flow, frequency analysis, error-level analysis, temporal tracking, and geometric consistency. FOVEA reports 97.39% accuracy and 0.99 AUROC on GenVideo, outperforming the strongest supervised baseline in the reported setting. Additional experiments examine watermark removal, video duration, cross-benchmark transfer, backbone sensitivity, and component ablations.

Critical strengths
  1. FOVEA combines high-level VLM reasoning with explicit computer-vision measurements, providing a more interpretable alternative to relying solely on unconstrained VLM judgments. Its structured evidence records also facilitate tracing the basis of individual predictions.
  2. The reported results are strong, and the component ablations suggest that CV-based analysis and adaptive tool routing contribute substantially to performance.
Critical weaknesses
  1. Incomplete validation of the core design choices. Although the paper provides component ablations, it does not sufficiently compare the proposed fusion and routing strategies against simpler alternatives, such as majority voting, fixed-weight aggregation, or running all analysts. Consequently, it remains unclear whether the reported gains depend on the specific proposed mechanisms or primarily result from combining multiple existing tools.

  2. Limited evidence for cross-generator generalization. The decision process relies partly on VLM-generated confidence estimates and thresholds calibrated using only 100 labeled videos. High-confidence correctness on this calibration set does not establish calibration on unseen generators or different backbones. In addition, although the framework records evidence chains, the correctness and causal relevance of the extracted evidence are not independently validated.

Questions for authors
  1. What specific algorithmic contribution distinguishes FOVEA's evidence-fusion and tool-routing mechanisms from existing tool-augmented reasoning frameworks and conventional ensemble methods? Can the authors provide direct comparisons against simpler fusion and routing strategies?
  2. How well do the confidence estimates remain calibrated on held-out videos from unseen generators? What evidence supports the correctness and forensic relevance of the intermediate evidence records?
Reviewer_MvfjOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

FOVEA reframes AI-generated video detection as forensic reasoning that outputs a verdict plus a grounded evidence chain, instantiated as a training-free neuro-symbolic pipeline. A VLM planner routes videos to five analysts (style, physics, spatial, temporal, watermark), each combining VLM reasoning with deterministic CV sub-skills, and a reliability-aware judge fuses the evidence. With Claude Opus 4.6 it reports 97.39% accuracy and 0.99 AUROC on GenVideo across ten unseen generators, beating the strongest supervised baseline by 5.93 points, with robustness to watermark removal and transfer to GenVidBench.

Critical strengths

The framing is clean and the engineering is thorough, with a planner, typed evidence contracts, and a three-stage judge. The paper is unusually candid about its own confounds, documenting the GenVideo duration bias, the watermark-driven early exits, the fact that evidence-chain correctness was not independently validated, and that the duration control is not a full deconfounding. The ablations isolating CV grounding and sub-skill routing (removing them costs 13 and 12.6 points) are informative, and the Pure VLM baseline at 42.92% shows the framework matters beyond the backbone.

Critical weaknesses
  • The core comparison is not sound as a like-for-like claim. Frozen frontier VLMs with undisclosed, web-scale pretraining are compared against detectors trained on a single generator (Pika) under the GenVideo protocol, and the paper concedes benchmark-video exposure cannot be ruled out.
  • The headline backbone (Claude Opus 4.6) reaches 100% on several generators, which is more consistent with memorization or watermark shortcuts than with transferable forensic reasoning, and the 89% watermark-driven early-exit rate reinforces this.
  • The central conceptual contribution, that evidence chains constitute forensic reasoning, is undercut by the authors’ own statement that the chains’ correctness and causal relevance are unvalidated, so the interpretability claim is traceability, not grounded reasoning.
  • Reproducibility is weak by construction, since results depend on proprietary API models that change over time, and the cost ($0.25 per video) and non-deterministic routing limit practical deployment.
  • The “training-free” label is also soft, as thresholds are hand-tuned on a 100-video calibration set with manual selection prioritizing a high-precision early exit.
Questions for authors
  • Given undisclosed VLM pretraining and conceded possible benchmark exposure, what evidence shows the gains are transferable forensic reasoning rather than memorization of GenVideo-era generators?
  • With watermark cues driving 89% of early exits and 100% accuracy on several generators, how do you separate watermark and content shortcuts from genuine artifact detection under a fully watermark-free, duration-matched protocol?
  • Since the evidence chains’ correctness is unvalidated, on what basis is this “reasoning” rather than post-hoc traceable logging?
  • How should a reader weigh a frozen frontier VLM against single-generator-trained baselines as a controlled comparison?
  • What is the variance of accuracy across repeated runs given non-deterministic planner and judge routing?
Reviewer_PR31OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The approach (FOVEA) puts together a VLM-based reasoning pipeline for various attributes of a video (style, physics, spatial, temporal, watermark presence) to detect real from synthetic videos. Notably, no task-specific training is performed. FOVEA seems to be a framework where one can drop-in replace other VLMs for a low-cost boost to performance. The authors show results mainly on GenVideo benchmark and consider somewhat latest baselines for comparisons.

Critical strengths

One of the most interesting strength of this work is to repurpose general-purpose LLMs for forensics. The reliability of this is still somewhat unclear--since calibration robustness isn't explored yet--but overall the work goes in the promising direction of adapting the performance of LLMs towards forensics.

Critical weaknesses
  1. Evaluating mainly on GenVideo and, to some extent, GenVideoBench, is not sufficient for a very critical reason: a. the GenVideo dataset (and its successor GenVidBench) contains watermarks with VARIABLE placement across screens: because these datasets compile unedited raw outputs directly from commercial video generators like Sora, Pika, and Kling, it inadvertently includes the visible digital overlays or provenance watermarks embedded by those proprietary platforms. As a result, the forensics cues from a VLM-based model are latching on to something very obvious. I looked through the watermark experiment where the authors remove the bottom right corner: however, the watermark presence is not guaranteed to be in the bottom examples. I would like to request the authors to evaluate on AIGVDbench and on SynthForensics which were developed to alleviate this major watermark-shortcut limitation of GenVideo and GenVideoBench,

  2. Every single table should report threshold-independent AUC as well. I only see it in Table 3 of supplementary. Please converge on a common set of evaluation metrics and please ensure that those are the ones you use.

  3. Calibration seems like a crucial step to make the scores of a VLM reliable. Please show the stability of the calibration step as a function of different calibration sets.

Questions for authors
  1. As mentioned in point no. 1: I would like to observe the performance on AIGVDbench and on SynthForensics, since these benchmarks are better at addressing the watermarking issue, which was a known limitation of GenVideo and GenvidBench.

  2. The models compared against in Table 1 of the main paper do not seem to be state of the art. Please consider benchmarks such as https://safe-video-2025.dsri.org to arrive at the methods you would compare against. Looks like the winning model, "waverep" is available -- https://github.com/grip-unina/WaveRep-SyntheticVideoDetection/tree/main.

  3. It is hard to gain clarity on performance since the evaluation metrics across tables are varying: please update the tables with a common set of metrics (adding AUC for threshold-independent eval).

  4. What is the standard deviation across varying calibration set?

  5. Since FOVEA as a framework can wrap around any model, I am curious to understand how does FOVEA perform around any state-of-the-art open-source VLM?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.