BusterX++: Towards Unified Cross-Modal AI-Generated Content Detection and Explanation with MLLM

arXiv:2507.14632 · cs.CV · Submitted 2025-07-19 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "BusterX++: Towards Unified Cross-Modal AI-Generated Content Detection and Explanation with MLLM".

Jane: Here's the gist of the paper. BusterX++ presents a unified MLLM for joint image and video detection with interpretable reasoning that demonstrates superior cross-modal capability transfer through a single-stage pure RL strategy The key insight reveals that SFT imposes lower policy entropy, which restricts the policy search space and dampens exploratory freedom, in contrast to single-stage pure RL which maintains higher policy entropy throughout training, effectively unlocking the spontaneous emergence of cross-modal capability transfer between image and video forensics Unified Benchmark Construction The paper introduces GenBuster-Bench++, a meticulously curated unified benchmark of AI-generated images and videos designed to observe and quantitate cross-modal capabilities This benchmark contains balanced image and video samples spanning recent generation models and diverse realworld scenarios, ensuring a clean, realistic, and difficultyaligned environment specifically engineered to observe and quantitate cross-modal capabilities.

Tom: The construction involves sourcing real images/videos from OpenVid-1M HD (Nan et al., 2024) and generating fake samples using a pipeline involving Reddit’s official API, Qwen-2.5-VL, and models like FLUX (Labs et al., 2025) and Seedance 1.0 (Gao et al., 2025).

Jane: A rigorous two-stage filtering pipeline is implemented for fake samples, where a synthetic sample is retained only when it successfully deceived at least 3 out of 5 experts into classifying it as real, ensuring the extreme deceptiveness and high difficulty of the dataset.

Tom: Training Strategy Analysis The paper argues that combining image and video data enhances unified learning through latent crossmodal complementarities, where high-quality static images provide rich, fine-grained texture and sharp boundary priors.

Title and authors: Tom: So to kick things off, Jane, let's talk about the full title of this paper, "BusterX++: Towards Unified Cross-Modal AI-Generated Content Detection and Explanation with MLLM." What does that actually mean for us?

Jane: It sounds pretty technical because it’s using an MLLM—a Multimodal Large Language Model—to do joint image and video detection, which is a bit of a mouthful.

Lu: Essentially, the paper is proposing this unified model as the central piece that handles both image and video data within one shared representation space for joint detection with interpretable reasoning.

Tom: So instead of having separate tools for image forensics and video forensics, this model tries to bridge that gap by using a single MLLM structure.

Jane: And they also introduce GenBusterBench++, which is their meticulously curated benchmark of AI-generated images and videos designed specifically to observe and quantify those cross-modal capabilities in a controlled setting.

Meng: A benchmark is only useful if it’s relevant; so I need to know how hard this test set actually is, especially since we're talking about AI-generated content.

Tom: The authors constructed this benchmark by sourcing real data from OpenVid-1M HD and then generating fake samples using pipelines involving models like FLUX and Seedance one point zero, covering both recent generation models and diverse real-world scenarios <ref:2507.14632#pg1,recent generation models and diverse>.

Jane: And they didn't just throw random stuff in there; they implemented a rigorous two-stage filtering pipeline for the fake samples to ensure extreme difficulty and high deception rates.

Lu: That filtering process is pretty strict; a synthetic sample only makes it in if it successfully deceives at least three out of five experts into classifying it as real, which sets a very high bar for the test set.

Tom: So they’re setting the bar high from the start by creating this controlled environment to see what our unified model can actually do.

Jane: Exactly, because that controlled setting is what lets them properly measure those cross-modal interactions in a way that’s hard to get elsewhere.

Meng: I'm still thinking about how much of that difficulty relies on the specific models used for generation; if they drop out of the test set, does it matter?

Tom: They explicitly mention covering video samples generated by thirteen state-of-the-art video generators and image samples from seven state-of-the-art image generators.

Lu: That diversity across different generators means the model isn't just overfitting to one generation pattern; it has to learn the underlying structural invariants.

Jane: So, they are testing if this unified approach can handle the sheer variety of what’s out there in AI media right now.

The paper's summary: Tom: Now, let’s get into what BusterX++ is actually doing in terms of its training strategy. It moves away from supervised fine-tuning and adopts a single-stage pure reinforcement learning approach to see the results.

Jane: They are skipping the supervised fine-tuning phase because they found that reliable chain of thought supervision is hard to get because visual forgery explanations aren't unique.

Lu: That’s the key insight: SFT imposes lower policy entropy, which restricts the policy search space and dampens exploratory freedom, whereas pure RL maintains higher policy entropy throughout training.

Tom: So pure RL keeps that exploration going, which they argue unlocks the spontaneous emergence of cross-modal capability transfer between image and video forensics.

Jane: That means they believe this method allows for a kind of emergent learning where the model develops those links between modalities on its own, rather than being explicitly taught them during fine-tuning.

Meng: But that sounds risky; if you don't guide it through SFT, how do we ensure it doesn't just wander into nonsensical outputs?

Tom: They use DAPO for optimization, which optimizes the policy model by maximizing an objective function that rewards outputs based on the reward model scores.

Jane: So they have a mechanism to guide the exploration without needing that explicit supervision during fine-tuning, using these reward model scores to steer the pure RL process.

Lu: And they show this training pipeline is designed to exploit complementary supervision across images and videos by leveraging latent cross-modal complementarities inherent in the data itself.

Tom: They suggest high-quality static images give rich texture and sharp boundary priors, which helps when dealing with frame-level inconsistencies in synthetic videos.

Jane: That’s a very concrete way of saying that the static image data provides physical structure that mitigates temporal noise found in video sequences.

The paper's improvements: Tom: Let's focus on the specific improvements they suggest they made to this system, because these are the mechanisms driving their results.

Lu: One key improvement is demonstrating cross-modal transfer: specifically, fine-grained texture and sharp boundary priors native to highresolution images can transfer over to mitigate frame-level inconsistencies in synthetic videos.

Jane: That’s a really strong claim because it means the model learns how to use the static image information as a corrective mechanism for video errors.

Tom: And conversely, they also showed image-to-video feature transfer, where the model bypasses deceptive global aesthetics in images to expose forgery by prioritizing frame-local artifacts.

Lu: So, instead of looking at the whole picture aesthetically, it focuses on those tiny local anomalies that are transferable to the video domain.

Meng: That sounds like a much more practical way to look for manipulation; focusing on local inconsistencies rather than trying to catch a global fake look.

Tom: And they also highlighted the model’s ability to leverage internal parametric world knowledge to cross-verify physical and semantic plausibility, which adds another layer of verification.

Jane: That’s smart because it means the model isn't just matching pixels; it’s checking if what it sees makes sense in terms of physics and context.

Lu: And that leads directly into their explanation quality validation, where they compared BusterX++ explanations against GPT-five point two using a strict double-blind user preference evaluation.

Tom: The results showed BusterX++ explanations were preferred over GPT-five point two in eighty-two percent of cases, which suggests the model’s reasoning is aligned with human expert preferences.

Jane: And they validated this by having experts confirm that specific visual artifacts cited in BusterX++’s reasoning blocks were factually present in the media in eighty-seven percent of sampled cases.

Conclusion: Tom: Alright, so to wrap up, the main implication here is that we have a unified MLLM for cross-modal AI-generated content detection with interpretable reasoning.

Jane: The key takeaway is that using a single-stage pure RL strategy drives the emergence of cross-modal capability transfer because it unlocks that exploration needed for this kind of synergy.

Lu: And they prove this transfer happens because high resolution image priors help mitigate video inconsistencies and dynamic video data helps enrich robustness against static images.

Meng: So, from a production perspective, the performance is solid compared to other MLLMs on GenBusterBench++, which suggests it’s a viable candidate for serious testing.

Tom: And we have strong evidence that its explanations are actually useful because they ground themselves in authentic visual evidence, which is a major win for trust.

Jane: It seems like this paper shows a solid path forward for making AI content detection tools more reliable by focusing on verifiable reasoning rather than just chasing high accuracy numbers.

Lu: The future work will likely involve tackling the challenge of adapting to the latest generative technologies, which means they need to keep enhancing the framework for continuous improvement.

Meng: I'm looking forward to seeing how they handle moving this from a lab setting into something that can actually run reliably under heavy load.

Tom: So, BusterX++ is showing us how pure RL can be a powerful tool for unifying complex multimodal tasks like visual forensics.

Jane: It’s definitely worth tuning in because it offers a concrete way to see how we can build AI systems that reason across different types of visual data.

Haiquan Wen, Tianxiao Li, Zhenglin Huang, Yiwei He, Guangliang Cheng

University of Liverpool

cs.CV

Submitted: 2025-07-19

Updated: 2026-11-13

Code: https://github.com/l8cv/BusterX

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 84/100

The gist: BusterX++ presents a unified MLLM for joint image and video detection with interpretable reasoning that demonstrates superior cross-modal capability transfer through a single-stage pure RL strategy…

Key concepts

GenBusterBench++
This is a specialized benchmark containing balanced, high-difficulty AI-generated images and videos. It was created to rigorously test the model's ability to detect fakes across different generation models and real-world scenarios, ensuring the testing environment is clean and challenging.
Single-Stage Pure RL Strategy
Instead of training through supervised fine-tuning first, this method uses direct Reinforcement Learning. This approach maintains high policy entropy during training, which is crucial for allowing the model to discover new ways to transfer knowledge between image and video detection spontaneously.
Cross-Modal Capability Transfer
This refers to the model's ability to leverage knowledge learned from one modality (like static images) and apply it effectively to another (like dynamic videos). The paper shows this happens because the model learns robust features, such as texture from images, that help it detect inconsistencies in video sequences.
Explanation Quality Validation
The model's reasoning is checked by human experts. BusterX++ explanations are preferred over other models because they are highly aligned with what human experts consider good reasoning. Furthermore, 87% of cited visual artifacts were confirmed to be real in the media, proving the explanations are factually grounded.

Terminology

Summary

BusterX++ presents a unified MLLM for joint image and video detection with interpretable reasoning that demonstrates superior cross-modal capability transfer through a single-stage pure RL strategy The key insight reveals that SFT imposes lower policy entropy, which restricts the policy search space and dampens exploratory freedom, in contrast to single-stage pure RL which maintains higher policy entropy throughout training, effectively unlocking the spontaneous emergence of cross-modal capability transfer between image and video forensics

Unified Benchmark Construction

The paper introduces GenBuster-Bench++, a meticulously curated unified benchmark of AI-generated images and videos designed to observe and quantitate cross-modal capabilities This benchmark contains balanced image and video samples spanning recent generation models and diverse realworld scenarios, ensuring a clean, realistic, and difficultyaligned environment specifically engineered to observe and quantitate cross-modal capabilities. The construction involves sourcing real images/videos from OpenVid-1M HD (Nan et al. p. 2

Where the paper says it

Real Images/Videos. We source a large number of real images and videos from OpenVid-1M HD (Nan et al.

, 2024) and generating fake samples using a pipeline involving Reddit’s official API, Qwen-2. p. 2
Where the paper says it

In addition, we construct the following pipeline for generation: we use Reddit’s official API to obtain social media images that cover a wide range of real-world scenarios. Then, we employ Qwen-2.

5-VL, and models like FLUX (Labs et al. p. 2
Where the paper says it

These captions then serve as prompts for various diffusion models to generate synthetic content. Finally, there are models such as FLUX (Labs et al.

, 2025) and Seedance 1. p. 2
Where the paper says it

Finally, there are models such as FLUX (Labs et al. , 2025) and GPT-4o (OpenAI, 2024) for image generation and Seedance 1.

0 (Gao et al. p. 2
Where the paper says it

, 2025) and GPT-4o (OpenAI, 2024) for image generation and Seedance 1. 0 (Gao et al.

, 2025). A rigorous two-stage filtering pipeline is implemented for fake samples, where a synthetic sample is retained only when it successfully deceived at least 3 out of 5 experts into classifying it as real, ensuring the extreme deceptiveness and high difficulty of the dataset. p. 3
Where the paper says it

In the second stage, these identified samples are re-examined to separate those that were actually synthetic. We implemented a strict majority-vote protocol: a synthetic sample was retained in our final benchmark only when it successfully deceived at least 3 out of 5 experts into classifying it as real, ensuring the extreme deceptiveness and high difficulty of the dataset.

Training Strategy Analysis

The paper argues that combining image and video data enhances unified learning through latent crossmodal complementarities, where high-quality static images provide rich, fine-grained texture and sharp boundary priors. Conversely, video sequences naturally encompass diverse compression noise and temporal variations, potentially enriching the model’s physical robustness when evaluating subtle artifacts in single isolated images. p. 4

Where the paper says it

Specifically, highquality static images provide rich, fine-grained texture and sharp boundary priors, which may conceptually assist the model in locating subtle, pixel-level forgery traces within video frames. Conversely, video sequences naturally encompass diverse compression noise and temporal variations, potentially enriching the model’s physical robustness when evaluating subtle artifacts in single isolated images.

The authors bypass Supervised Fine-Tuning (SFT) to unlock latent cross-modal synergies because reliable CoT supervision is difficult to obtain due to the non-unique nature of visual forgery explanations. p. 3
Where the paper says it

However, applying this paradigm to AI-generated content detection introduces unique challenges. Reliable CoT supervision is difficult to obtain because explanations for visual forgery detection are often non-unique and depend on subtle visual cues such as textures, lighting, motion artifacts, and local inconsistencies.

The proposed training pipeline uses DAPO, which optimizes the policy model by maximizing an objective function that rewards outputs based on the reward model scores. p. 0
Where the paper says it

Notably, our empirical results demonstrate that a single-stage, pure RL strategy driven strictly by simple rewards without any SFT warm-up or complex reward shaping consistently yields competitive or superior performance compared to the strong SFT+RL pipeline across both single-modality and unified tasks. To better understand this behavior, we analyze training dynamics and observe lower policy entropy under SFT initialization, which dampens the model’s exploratory freedom.

Performance and Comparison

Extensive experiments demonstrate that BusterX++ achieves state-of-the-art performance, highlighting the powerful potential of RL for unified cross-modal visual reasoning. Table 1 shows that BusterX++ achieves notable overall accuracy on GenBuster-Bench++, outperforming other MLLMs such as Qwen3. p. 4

Where the paper says it

6 - ✓ ✓ ✓ 85. 1 Performance on GenBuster-Bench++ As shown in Table 1, BusterX++ achieves notable overall accuracy.

5-122B-A10B and Qwen3. p. 4
Where the paper says it

6V-Flash 96. 5-122B-A10B 82.

6-27B across both image and video tasks. p. 4
Where the paper says it

Conversely, video sequences naturally encompass diverse compression noise and temporal variations, potentially enriching the model’s physical robustness when evaluating subtle artifacts in single isolated images. By jointly training on both modalities within a shared optimization space, the model is expected to leverage such complementary information, thereby achieving stronger mutual generalization across both image and video forensics tasks.

Furthermore, Table 4 confirms that the single-stage pure RL strategy consistently outperforms SFT baselines, as the pure RL policy steadily catches up with and eventually outshines the strong SFT+RL pipeline in the later phase. p. 5
Where the paper says it

Although the SFT baseline inherently starts with a higher accuracy reward, our single-stage RL approach exhibits a remarkably steep optimization curve. Driven strictly by sparse accuracy rewards without any SFT warmup, the pure RL policy steadily catches up with and eventually outshines the strong SFT+RL pipeline in the later phase.

Cross-Modal Transfer Mechanisms

The analysis of training dynamics reveals two symmetric emergent behaviors when skipping SFT and preserving high strategy entropy. First, Video-to-Image Robustness Transfer occurs because extensive training on dynamic video sequences equips the pure RL model with an intrinsic tolerance to temporal ambiguities, which directly benefits static image detection. p. 5

Where the paper says it

Our empirical analysis reveals two symmetric emergent behaviors that manifest when skipping SFT and preserving high strategy entropy: 1. Video-to-Image Robustness Transfer: Conversely, extensive training on dynamic video sequences equips the pure RL model with an intrinsic tolerance to temporal ambiguities, which directly benefits static image detec6

Second, Image-to-Video Feature Transfer shows that the model bypasses deceptive global aesthetic to expose forgery by prioritizing frame-local artifacts, enabling the transfer of image-level sensitivity to the video domain. p. 7
Where the paper says it

Fine-grained Visual Perception: The pure RL paradigm grants the model a heightened sensitivity toward micro-level, low-level details that are easily overlooked by traditional SFT-initialized pipelines. For instance, in the image-to-video transfer scenario (Figure 4b), BusterX++ bypasses the deceptive global aesthetic to expose the forgery, showing robust perception of fine-grained artifact invariants.

The model also demonstrates an elegant capacity for leveraging internal parametric world knowledge to cross-verify physical and semantic plausibility. p. 7
Where the paper says it

For instance, in the image-to-video transfer scenario (Figure 4b), BusterX++ bypasses the deceptive global aesthetic to expose the forgery, showing robust perception of fine-grained artifact invariants. World Knowledge: Beyond tracking statistical pixel-level footprint anomalies, BusterX++ demonstrates an elegant capacity for leveraging internal parametric world knowledge to cross-verify physical and semantic plausibility.

Explanation Quality Validation

A strict double-blind user preference evaluation compared BusterX++ explanations against GPT-5.2 showed that BusterX++ explanations are preferred over GPT-5. p. 7

Where the paper says it

Each expert cast an independent vote evaluating multiple analytical dimensions, including evidence grounding, specificity, and completeness, with the winning explanation determined by a strict 3-out-of-5 majority consensus. We report that BusterX++ explanations are preferred over GPT-5.

2 explanations in 82% of cases, indicating that the explanation quality is highly aligned with human expert preferences. p. 7
Where the paper says it

We report that BusterX++ explanations are preferred over GPT-5. 2 explanations in 82% of cases, indicating that the explanation quality is highly aligned with human expert preferences.

To verify the authenticity of its reasoning, experts confirmed that specific visual artifacts cited in BusterX++’s blocks were factually present in the media in 87% of sampled cases, strongly validating that single-stage pure RL forces the policy to anchor its decisions on authentic visual evidence. p. 7
Where the paper says it

, texture blending, reflection anomalies) were factually present in the media in 87% of the sampled cases. This high grounding rate strongly validates that single-stage pure RL forces the policy to anchor its decisions on authentic visual evidence.

Robustness and Generalization

BusterX++ shows remarkable stability against common visual perturbations, including JPEG Compression (quality=70) and Gaussian Noise (σ=5), even without specific training on degraded data. Table 4 illustrates this robustness, where BusterX++ maintains solid performance under the harshest degradation cascade. p. 6

Where the paper says it

While Gaussian Blur appears to have only a marginal effect on overall accuracy, it produces a mild decline within sub-categories. BusterX++ still delivers solid performance under the harshest degradation cascade, highlights its robustness and practical value in diverse and unpredictable real-world settings.

The model’s ability to adopt a structured, step-by-step analytical pattern across all scenarios demonstrates exceptional reasoning stability regardless of the task modality. p. 6
Where the paper says it

Due to space constraints, additional qualitative examples and case studies are presented in Appendix F, further demonstrating the effectiveness of BusterX++ across diverse scenarios. Stable Reasoning: Across all evaluated scenarios, BusterX++ consistently adopts a structured, step-by-step analytical pattern instead of producing a superficial, binary judgment.

This fine-grained visual perception allows the pure RL paradigm to bypass deceptive global aesthetics and expose forgery through localized anomalies. p. 7
Where the paper says it

Fine-grained Visual Perception: The pure RL paradigm grants the model a heightened sensitivity toward micro-level, low-level details that are easily overlooked by traditional SFT-initialized pipelines. For instance, in the image-to-video transfer scenario (Figure 4b), BusterX++ bypasses the deceptive global aesthetic to expose the forgery, showing robust perception of fine-grained artifact invariants.

The model also leverages deep context familiarity to verify the authentic coexistence of official Real Madrid kit markers alongside correct sponsor typography, showcasing world knowledge utilization. p. 7
Where the paper says it

World Knowledge: Beyond tracking statistical pixel-level footprint anomalies, BusterX++ demonstrates an elegant capacity for leveraging internal parametric world knowledge to cross-verify physical and semantic plausibility. In Figure 4a), it utilizes deep context familiarity to verify the authentic coexistence of official Real Madrid kit markers alongside correct sponsor typography.

Societal Impact and Limitations

AI-generated content detection models are not infallible, and errors could lead to false positives or false negatives, necessitating adoption alongside human oversight. The limitations include the challenge of adapting to the latest generative technologies present in GenBuster-Bench++, indicating a need for continuous framework enhancement. p. 7

Where the paper says it

We observe that samples generated with the latest technologies in GenBuster-Bench++ present a more significant challenge. This indicates that the rapid evolution of generative methods necessitates continuous adaptation and enhancement of our detection framework to maintain its efficacy.

Additionally, the authors suggest that post-training may be approaching a performance bottleneck, suggesting future work should explore other training stages. p. 7
Where the paper says it

Posttraining phase may be approaching a performance bottleneck in our task. To further advance the capabilities of MLLMs, future work should explore other training stages.

The model’s structured reasoning trajectory shows exceptional stability across all evaluated scenarios. p. 7
Where the paper says it

As detailed in Figure 4, the model systematically structures its <think> trajectory. It explicitly enumerates localized elements, showing exceptional reasoning stability regardless of the task modality.

Experimental Details

The experimental setup utilizes Qwen3.5-4B as the base model, employing AdamW with a learning rate of 1 × 10−6 for stable RL training. p. 4

Where the paper says it

To optimize processing efficiency while preserving structural features, we enforce a strict token-level computational budget of 64 tokens per frame/image. We employ AdamW (Loshchilov and Hutter, 2019) with a learning rate of 1 × 10−6 , which aligns with established community best practices for stable RL training.

For video processing, frames are sampled uniformly at a rate of 2. p. 4
Where the paper says it

5-4B (Qwen Team, 2026) as our base model. For video-level forensic processing, we sample frames uniformly at a rate of 2.

0 FPS, and a strict token-level computational budget of 64 tokens per frame/image is enforced. p. 4
Where the paper says it

For video-level forensic processing, we sample frames uniformly at a rate of 2. To optimize processing efficiency while preserving structural features, we enforce a strict token-level computational budget of 64 tokens per frame/image.

The training set GenBuster-Unified comprises 10,000 samples sourced from So-Fake-Set (Huang et al. p. 4
Where the paper says it

To construct a computationally efficient yet modality-balanced training environment, we introduce GenBuster-Unified, a lightweight unified training set comprising 10,000 samples in total. The instances are sourced and curated from So-Fake-Set (Huang et al.

, 2025b) and GenBuster200K (Wen et al. p. 4
Where the paper says it

The instances are sourced and curated from So-Fake-Set (Huang et al. , 2025b) and GenBuster200K (Wen et al.

, 25a). The DAPO RL strategy optimizes the policy model by maximizing the objective function defined in equation (2), which involves minimizing the expected reward across generated outputs. p. 4
Where the paper says it

By jointly training on both modalities within a shared optimization space, the model is expected to leverage such complementary information, thereby achieving stronger mutual generalization across both image and video forensics tasks. We adopt DAPO (Yu et al.

The prompt design carefully guides the model to analyze inconsistencies and provide a single-letter conclusion for both image and video inputs. p. 5
Where the paper says it

, 2025b) and GenBuster-200K (Wen et al. To ensure the synthetic data possesses both rigorous forensic depth and precise structural alignment, we feed these images and video frames into the teacher model along with a carefully engineered prompt.

Conclusion

In summary, BusterX++ utilizes a single-stage RL strategy optimized directly via sparse accuracy rewards for cross-modal AI-generated content detection. The results indicate that skipping the supervised finetuning phase and deploying direct RL maintains higher strategy entropy during post-training, which facilitates effective cross-modal knowledge transfer between the image and video domains. p. 7

Where the paper says it

To facilitate evaluation, we introduce GenBusterBench++, a curated unified benchmark consisting of 4,000 image and video samples. Our experimental results indicate that skipping the supervised finetuning phase and deploying direct RL maintains higher strategy entropy during post-training, which facilitates effective cross-modal knowledge transfer between the image and video domains.

Comprehensive evaluations demonstrate that this approach yields superior or competitive performance compared to strong SFT-initialized baselines, highlighting the potential of RL for cross-modal reasoning. p. 7
Where the paper says it

Our experimental results indicate that skipping the supervised finetuning phase and deploying direct RL maintains higher strategy entropy during post-training, which facilitates effective cross-modal knowledge transfer between the image and video domains. Comprehensive evaluations demonstrate that our approach yields superior or competitive performance compared to strong SFT-initialized baselines, highlighting the potential of RL for cross-modal reasoning.

The model’s explanations are preferred by human experts, and its grounding rate confirms that it anchors decisions on authentic visual evidence. p. 7
Where the paper says it

, texture blending, reflection anomalies) were factually present in the media in 87% of the sampled cases. This high grounding rate strongly validates that single-stage pure RL forces the policy to anchor its decisions on authentic visual evidence.

Appendix Details

The appendix provides comprehensive details regarding training configurations, evaluation parameters, and hyperparameter selections utilized in the study to ensure full reproducibility. The prompt design for both video and image inputs is carefully engineered to guide the model toward a structured analytical pattern rather than superficial judgment. p. 6

Where the paper says it

Due to space constraints, additional qualitative examples and case studies are presented in Appendix F, further demonstrating the effectiveness of BusterX++ across diverse scenarios. Stable Reasoning: Across all evaluated scenarios, BusterX++ consistently adopts a structured, step-by-step analytical pattern instead of producing a superficial, binary judgment.

The user preference evaluation involves blind presentation of media and two candidate explanations, with experts judging criteria such as evidence grounding, specificity, internal consistency, completeness, and clarity. p. 7
Where the paper says it

The evaluation was conducted by the exact same panel of 5 forensic experts who filtered our benchmark. For each sample, experts were blindly presented with the media and two anonymized candidate explanations (Ours vs.

The final responses are provided in the appendix for full verification of the model's reasoning chains.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Unified Cross-Modal Detection Architecture: Implement BusterX++ to create a single MLLM that jointly processes image and video data within a shared representation space, allowing for joint image and video detection with interpretable reasoning.

  2. Training Strategy Optimization: Adopt a single-stage pure RL strategy driven strictly by sparse outcome rewards instead of the traditional SFT+RL paradigm to maintain higher policy entropy throughout training, which is shown to unlock the spontaneous emergence of cross-modal capability transfer between image and video forensics.

  3. Benchmark Integration for Control: Utilize GenBusterBench++ as a meticulously curated, difficulty-aligned benchmark containing balanced image and video samples spanning recent generation models and diverse realworld scenarios to provide a controlled setting for investigating cross-modal interactions.

  4. Cross-Modal Knowledge Transfer Capabilities: The improved system will exhibit transferable reasoning, specifically demonstrating that fine-grained texture and sharp boundary priors native to highresolution images can transfer to mitigate framelevel inconsistencies in synthetic videos.

  5. Robustness Against Perturbations: Enhance the model's resilience by training it on degraded data, as BusterX++ shows remarkable stability against these low-level distortions, even without specific training on degraded data, including JPEG Compression and Gaussian Noise.

  6. Explainability and Grounding: Leverage the DAPO RL strategy to generate structured explanations where intermediate reasoning chains are factually present in the media in 87 percent of cases, ensuring that single-stage pure RL forces the policy to anchor its decisions on authentic visual evidence.

Abstract

The rapid advancement of generative AI has substantially improved image and video synthesis, amplifying the risk of multimodal visual misinformation. Recent MLLMs have shown promise for transparent AI-generated content detection through reasoning and explanation, yet existing approaches largely treat image and video forensics as isolated tasks, leaving cross-modal synergies underexplored. To address this, we present BusterX++, a unified MLLM for joint image and video detection with interpretable reasoning. We also introduce GenBuster-Bench++, a curated benchmark containing balanced image and video samples spanning recent generation models and diverse real-world scenarios, with synthetic images and videos screened using the same human difficulty criterion. Using this controlled setting, we revisit the widely adopted SFT+RL post-training paradigm. Notably, in the unified image-video setting, our results demonstrate that a single-stage pure RL strategy driven primarily by sparse outcome rewards outperforms a strong SFT+RL baseline. We observe that single-stage pure RL maintains higher policy entropy than SFT-initialized RL throughout training. This observation suggests a possible role for exploration in the performance gains observed under unified image-video training. Extensive experiments demonstrate that BusterX++ achieves state-of-the-art performance, highlighting the powerful potential of RL for unified cross-modal visual reasoning.

Sources

Related papers