Technical Summary: BusterX++

Problem Statement

The rapid advancement of generative AI has significantly enhanced the quality of image and video synthesis, thereby amplifying the risk of multimodal visual misinformation. While recent Multimodal Large Language Models (MLLMs) show promise in detecting AI-generated content (AIGC) through transparent reasoning and explanation, existing approaches largely treat image and video forensics as isolated tasks. This separation leaves cross-modal synergies underexplored. Furthermore, the prevailing training paradigm for reasoning-oriented MLLMs involves a two-stage process: Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL). The authors question whether this SFT initialization, which relies on curated demonstrations, inadvertently constrains the model's policy search space, dampening the exploratory freedom necessary to discover subtle, non-unique visual forgery cues and hindering the spontaneous emergence of cross-modal capability transfer.

Methodology

1. GenBuster-Bench++: A Unified Benchmark

To address the lack of controlled evaluation environments for cross-modal interaction, the authors introduce GenBuster-Bench++.

• Composition: A meticulously curated dataset containing 4,000 samples (2,000 images and 2,000 videos).
• Sources: Real samples are sourced from OpenVid-1M HD. Synthetic samples are generated using 13 state-of-the-art (SOTA) video generators (e.g., Runway Gen-3, Sora, Kling 2.0) and 7 SOTA image generators (e.g., FLUX, GPT-4o, Ideogram).
• Curation: The dataset undergoes a rigorous two-stage human filtering pipeline. Crucially, synthetic samples are retained only if they successfully deceive at least 3 out of 5 forensic experts, ensuring high difficulty and realism.
• Alignment: The benchmark ensures balanced difficulty across modalities and diverse real-world scenarios (9 categories), eliminating modality-difficulty imbalances and domain divergences found in previous benchmarks.

2. BusterX++: Single-Stage Pure RL Training

The core methodological contribution is BusterX++, a unified MLLM trained via a streamlined, single-stage Reinforcement Learning (RL) pipeline that completely bypasses SFT.

• Training Strategy: The model is trained directly on a unified dataset (GenBuster-Unified, 10,000 samples) using the DAPO (Direct Policy Optimization) algorithm.
• Reward Function: The optimization is driven strictly by sparse outcome rewards:
  • Format Reward: Penalizes deviation from the required output structure.
  • Accuracy Reward: Awards +1 for correct classification (Real/Fake) and 0 otherwise.
  • Soft Overlong Penalty: Penalizes excessive response length.
• Rationale for Skipping SFT: The authors hypothesize that SFT imposes lower policy entropy, restricting the model to reproducing textual reasoning patterns rather than exploring discriminative visual evidence. By removing the SFT warm-up, the pure RL strategy maintains higher policy entropy, encouraging the model to explore the solution space and discover latent cross-modal synergies.

3. Unified Cross-Modal Learning

The training data combines image and video modalities within a shared optimization space. The authors posit that:

• Image-to-Video Transfer: Fine-grained texture and sharp boundary priors from high-resolution images help mitigate frame-level inconsistencies in synthetic videos.
• Video-to-Image Transfer: Dynamic environmental variations and temporal inconsistencies from video sequences enrich the model's robustness when auditing static images, particularly against motion blur or lighting artifacts.

Key Contributions

1. GenBuster-Bench++: A difficulty-aligned, unified benchmark for AI-generated image and video detection, featuring balanced samples from diverse SOTA generators and rigorous human curation.
2. BusterX++: A unified MLLM architecture that achieves state-of-the-art performance in both image and video forensics through a single-stage, pure RL training paradigm.
3. Paradigm Shift in Training: Empirical evidence demonstrating that a single-stage pure RL strategy driven by sparse rewards outperforms the conventional SFT→RL pipeline. The authors identify that SFT reduces policy entropy, limiting the model's ability to explore and transfer capabilities across modalities.
4. Cross-Modal Synergy: The demonstration that unified training enables spontaneous capability transfer, where image-level priors improve video detection and vice versa, leading to superior generalization.

Experimental Results

• Performance on GenBuster-Bench++: BusterX++ achieves an overall accuracy of 81.8%, significantly outperforming the strong SFT+RL baseline (77.8%) and other open/closed-source MLLMs (e.g., Qwen3.5-122B, GPT-5.2).
• Ablation Studies:
  • Training Pipeline: Pure RL consistently outperforms SFT-initialized RL across single-modality and unified settings.
  • Modality: Unified training (images + videos) yields higher accuracy than single-modality training, even within the native modality, confirming the benefit of cross-modal complementarity.
• Robustness: The model demonstrates remarkable stability against common visual perturbations (JPEG compression, Gaussian noise, Gaussian blur) and degradation cascades, maintaining high accuracy without specific training on degraded data.
• Case Studies:
  • Video-to-Image: Pure RL correctly identifies motion blur in real sports photos as natural, whereas the SFT baseline misinterprets it as digital compositing artifacts.
  • Image-to-Video: Pure RL detects localized frame-level artifacts in synthetic videos that the SFT baseline misses due to over-reliance on global temporal consistency.
• User Preference: In a double-blind evaluation with forensic experts, BusterX++ explanations were preferred over GPT-5.2 in 82% of cases. Furthermore, 87% of the visual artifacts cited in BusterX++'s reasoning were factually grounded in the media.

Significance and Claims

The paper claims that unified visual forensics benefits not only from architectural unification but also from training strategies that encourage transferable reasoning across modalities. The primary significance lies in the finding that SFT can be detrimental to the discovery of cross-modal capabilities in AIGC detection. By maintaining high policy entropy through a pure RL approach, the model is able to spontaneously learn to transfer fine-grained visual priors between images and videos.

The authors conclude that BusterX++ highlights the powerful potential of RL for unified cross-modal visual reasoning, offering a more effective alternative to the standard SFT→RL paradigm for detecting and explaining multimodal visual misinformation. They acknowledge limitations regarding the rapid evolution of generative technologies, noting that continuous adaptation of the framework will be necessary to maintain efficacy against future models.