Research map / Video / Video-language understanding

BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation

NeurIPS 2026 · Video › Video-language understanding · approach: Video LLMs with fine-grained understanding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces GenBuster-200K, a fair high-quality AI-generated video dataset, GenBuster-Bench with ID/OOD/Wild tracks and an MLLM-as-judge rationale score, and BusterX, a Qwen3.5-4B model trained with DAPO-style RL that treats reasoning as the detector.

Problem

Detecting AI-generated videos with MLLMs is limited by outdated datasets, aggregated binary evaluation, and black-box classification without forensic reasoning.

Approach

Formulate video forgery detection as RL-trained visual reasoning where the generated chain-of-thought serves as the detector, evaluated on a progressive benchmark (ID, OOD unseen commercial generators, in-the-wild) with a judge-scored rationale metric.

Key results

• BusterX reaches 86.7% ID accuracy and 87.8% Wild fake accuracy-based macro ACC of 77.5% in the Wild track, with 84.7 rationale score, the highest among compared MLLMs.
• On the OOD track BusterX achieves 87.8% accuracy, above Claude-Sonnet-4.6 (83.9% fake, 72.8% ACC overall per its table entries) and all classical detectors.
• Within the Qwen3.5 family, Wild accuracy rises steadily from 4B to 27B to 122B, though the 4B-based BusterX outperforms all of them.

Surprising findings

• Claude-Sonnet-4.6 achieves strong OOD accuracy yet only 61.6% on real ID videos, indicating an over-forgery bias.
• Generic MLLMs and classical detectors collapse on in-the-wild social media videos (e.g., Claude-Haiku-4.5 at 19.3% fake accuracy).

Limitations

• Limited to the visual modality; audio and audiovisual inconsistencies are not addressed.
• Explanation evaluation relies on an automated judge (Gemini 3 Flash Preview) and only on correctly predicted Wild samples.

Experimental setup

Datasets: GenBuster-200K OpenVid-1M

Benchmarks: GenBuster-Bench ID GenBuster-Bench OOD GenBuster-Bench Wild

Metrics: accuracy real accuracy fake accuracy rationale score (1-5 normalized) Spearman correlation

Baselines: 3D ResNeXt ViViT VideoMAE DeMamba GPT-5.2 Grok-4.1-fast Claude-Haiku-4.5 Claude-Sonnet-4.6 Qwen3.5-4B Qwen3.5-27B Qwen3.5-122B-A10B VideoVeritas

Builds on: DAPO GRPO DeepSeek-R1 OpenVid-1M

Closest papers by meaning

• Seeing What Matters: Generalizable AI-generated Video Detection with Forensic-Oriented Augmentation (NeurIPS 2025 · similarity 0.88)
• Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models (CVPR 2025 · similarity 0.88)
• Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning (ICLR 2026 · similarity 0.87)
• On Learning Multi-Modal Forgery Representation for Diffusion Generated Video Detection (NeurIPS 2024 · similarity 0.87)
• Identity-Aware Vision-Language Model for Explainable Face Forgery Detection (AAAI 2026 · similarity 0.86)
• THEMIS: Towards Holistic Evaluation of MLLMs for Scientific Paper Fraud Forensics (ICLR 2026 · similarity 0.86)