BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation".
Jane: Here's the gist of the paper. As generative video models become increasingly realistic, detecting AI-generated videos requires systems that offer both accuracy and interpretability The final sentence is BusterX outperforms several leading MLLMs (e.g., Qwen3.5, Claude-Sonnet-4.6) in both detection accuracy and rationale quality Dataset Introduction The paper introduces GenBuster-200K, a fair dataset of over 200,000 high-quality videos sourced from state-of-the-art generators This dataset is designed to address limitations in prior work by leveraging the latest generators and enforcing rigorous fairness constraints such as balanced gender and ethnicity distributions The construction pipeline involves three stages: Diverse Seed Generation, Description Generation, and Post-filtering Benchmark Design GenBuster-Bench is proposed as a diagnostic benchmark that decomposes the video forgery detection challenge along two critical axes: domain shifts and generational shifts This dual-axis design partitions the evaluation into three progressive tracks: In-Domain (Era: 2024), Out-of-Domain (Era: 2025), and In-the-Wild (Era: 2026) The evaluation protocol includes an MLLM-as-a-Judge protocol to assess the quality of the generated forensic explanations Model Training Strategy BusterX is developed as an MLLM baseline with Reinforcement Learning (RL) training, formulating detection as a visual reasoning task where the generated reasoning chain serves as the detector itself The RL training utilizes DAPO over GRPO because DAPO effectively prevents entropy collapse inherent to standard GRPO The total reward R is a weighted sum of three components: rformat, raccuracy, and rsoft overlong penalty Performance Results Experimental results demonstrate that BusterX outperforms several leading MLLMs (e.g., Qwen3.5, Claude-Sonnet-4.6) in both detection accuracy and rationale quality BusterX achieves the highest overall accuracy among all open-source MLLMs and classical detectors on the In-Domain (ID) track Furthermore, BusterX maintains strong robustness against real-world social media degradations in the Wild track The model’s full responses in Case Study shows that it successfully identifies physical inconsistencies like inconsistent wave progression Explainability Assessment BusterX achieves the highest Wild Rationale score as evaluated by the Gemini Judge on correctly predicted fake samples Human evaluation of the rationales generated by BusterX on the Wild track yielded a mean human score of 78.7, which shows a high level of forensic alignment with automated judges The model's ability to pinpoint subtle spatiotemporal inconsistencies is highlighted as a key capability Future Directions The current framework is limited to the visual modality, leaving audio manipulation and audiovisual semantic inconsistencies unaddressed Future work will extend toward audio-visual multimodal detection A societal impact note acknowledges the dual-use nature of this research and encourages access controls when deploying forensic tools in adversarial settings The paper also notes a scaling law within the Qwen3.5 family, indicating that larger parameter sizes bring consistent benefits in handling highly diverse, real-world forgeries The study concludes by emphasizing that formulating detection as a step-by-step visual reasoning task is essential The authors also detail the prompt design pipeline used to guide the model's response generation The paper references numerous related works in AI-generated video detection methods and MLLMs The paper concludes by addressing critical vulnerabilities in AI-generated video detection through three contributions The authors also provide detailed experimental details regarding training configurations and evaluation parameters for reproducibility The paper includes a section on related work covering early datasets, detection methods, and MLLMs The paper concludes by providing a list of references spanning various foundational and recent works in the field Page 1 --- BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation Forgery Detection and Explanation Haiquan Wen1∗ Yiwei He1∗ Zhenglin Huang1∗ Tianxiao Li1 Zihan Yu1 Xingru Huang5 Lu Qi4 Baoyuan Wu3 Xiangtai Li2 Guangliang Cheng1† Page 2 --- Table 1: Comparison with existing datasets.
Tom: GenBuster outperforms prior works in scale, visual quality and fairness (balanced gender, age, and ethnicity distributions).
Jane: In-the-wild refers to whether the dataset includes videos sourced from social media out-of-domain models or breaks down under real-world social media degradations Crucially, they also ignore the explainability of detection models In forensic scenarios, providing a human-readable rationale is just as important as the binary verdict itself Page 3 --- Lack of Forensic Reasoning: Recent methods that adapt MLLMs for detection either use them as auxiliary to assist binary classification, or merely to generate post-hoc explanations Neither approach treats step-by-step physical and spatiotemporal reasoning as the detection mechanism itself Page 4 --- To address these limitations, we introduce three contributions: 1) GenBuster-200K: A High-Quality and Fair Dataset 2) GenBuster-Bench: A Progressive Diagnostic Benchmark 3) BusterX: A Strong MLLM Baseline with RL Training Page 5 --- In-Domain (ID) Performance and Bias Mitigation.
Tom: On the ID track, BusterX achieves the highest overall accuracy among all open-source MLLMs and classical detectors For instance, Qwen3.5-122B exhibits a conservative bias (79.4% on Real, but 54.7% on Fake) Conversely, Claude-Sonnet-4.6 suffers from an over-forgery bias, which ruins its performance on real videos (61.6% Real ACC) In contrast, BusterX maintains a balanced performance, demonstrating that our training paradigm successfully aligns the model’s forensic perception without compromising its judgment on authentic videos Page 6 --- Out-of-Domain (OOD) Generalization.
Paper summary: Jane: So, to wrap up this discussion on BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation, we have seen how they built a comprehensive system involving a new dataset of over two hundred thousand videos, a progressive benchmark testing domain shifts across three tracks, and their model trained with reinforcement learning to make the detection process itself the reasoning chain.
Lu: What stands out is their focus on creating high-quality forensic explanations, as evidenced by the high Wild Rationale scores and human evaluation scores of seventy-eight point seven <ref:2505.12620#pg2>. This proves that generating a step-by-step visual reasoning task is a powerful way to move beyond simple binary verdicts in video forensics.
Tom: And when we look at the overall performance metrics, BusterX outperforms several leading MLLMs like Qwen3 point 5 and Claude-Sonnet-four point six in both detection accuracy and rationale quality <ref:2505.12620#pg1> <ref:2505.12620#pg1>. This suggests that their training paradigm, using DAPO over GRPO with a weighted reward function including accuracy and a soft overlong penalty, is effectively aligning the model's forensic perception without hurting its judgment on real videos.
Meng: The big picture for deployment is that while this work shows strong robustness against real-world social media degradations in the Wild track, it’s not yet ready for full audio-visual detection because they haven't tackled those other modalities yet <ref:2505.12620#pg1>. It’s a solid visual reasoning foundation, though.
Jane: So, the authors of BusterX are essentially showing that by treating detection as a visual reasoning task guided by careful dataset construction and reinforcement learning training, we can build systems that are both highly accurate at detecting fakes and provide detailed explanations for those detections <ref:2505.12620#pg1>. That’s what this paper is all about.
Conclusion: Tom: So we're wrapping up this look at BusterX, and it’s really worth focusing on what that title says: "BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation."
Jane: I think that title tells you exactly what the paper is about. It’s not just saying it finds fakes; it’s claiming to do both detection and explanation, which is a big deal in this field.
Lu: Exactly. The core idea here isn't just getting a yes or no on whether a video is fake. They are building a system where the model’s reasoning chain *is* the detector itself, so you get that explanation along with the verdict.
Meng: But how does that actually work without making things too slow? I mean, if we have to generate this whole step-by-step reasoning for every single video, what’s the computational cost when we try to deploy this at scale?
Tom: That’s where the training strategy comes in. They used reinforcement learning with something called DAPO over GRPO. It sounds complicated, but basically, they trained the model by rewarding it for doing things right—for being accurate and for not giving overly long, useless answers.
Jane: And those results are really compelling because BusterX beat some of the biggest models out there on both finding fakes and giving good explanations.
Lu: It's interesting because they showed that this approach works well across different challenges, like videos from the same time period versus real-world social media messiness. They found it maintains strong performance even when things get degraded in the wild.
Meng: Robustness is one thing, but accuracy on the hardest data is another. If we’re talking about what this means for production, can we rely on this kind of reasoning chain to catch subtle things that generic models miss?
Tom: The scaling law they mentioned with Qwen3 point five suggests that bigger parameter sizes bring consistent benefits when dealing with really diverse fakes, which is good news for future hardware planning <ref:2505.12620#pg1>.
Jane: So, BusterX shows us a direction where the model doesn't just guess; it thinks its way through the visual details to find the inconsistency and then explains exactly why it flagged it.
Lu: And that emphasis on step-by-step reasoning is crucial because it’s what makes the detection method itself more transparent and trustworthy.
Meng: For me, the next big question is whether this deep reasoning structure can be adapted to handle audio too, since they only focused on the visual side for now.
Tom: That's definitely a future direction they flagged—audio-visual detection is next on the list for them.
Haiquan Wen, Yiwei He, Zhenglin Huang, Tianxiao Li, Zihan Yu, Xingru Huang, Lu Qi, Baoyuan Wu
University of Liverpool University of Liverpool Nanyang Technological University The Chinese University of Hong Kong Shenzhen Wuhan University Hangzhou Dianzi University
cs.CV
Submitted: 2025-05-19
Updated: 2026-11-13
Code: https://github.com/l8cv/BusterX
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 97/100
The gist: As generative video models become increasingly realistic, detecting AI-generated videos requires systems that offer both accuracy and interpretability The final sentence is BusterX outperforms…
Key concepts
- GenBuster-200K
- This is a large, high-quality dataset containing over 200,000 videos created by state-of-the-art AI generators. It is specifically designed to be fair by ensuring balanced distributions of gender and ethnicity, addressing limitations in previous datasets.
- GenBuster-Bench
- This is a testing framework that evaluates video forgery detection along two axes: domain shifts (testing performance across different eras/sources) and generational shifts. It uses an MLLM-as-a-Judge protocol to assess the quality of the model's generated explanations.
- BusterX Training Strategy
- BusterX is trained using Reinforcement Learning (RL), treating detection as a visual reasoning task where the model generates its own reasoning chain. It uses DAPO over GRPO and optimizes a reward function that balances format adherence, accuracy, and response length.
- Visual Reasoning Task
- The core idea is to frame video forgery detection not just as classification but as a step-by-step visual reasoning process. The model's ability to articulate the physical inconsistencies it observes in the frames serves as its primary mechanism for making a forensic decision.
Terminology
Summary
As generative video models become increasingly realistic, detecting AI-generated videos requires systems that offer both accuracy and interpretability The final sentence is BusterX outperforms several leading MLLMs (e. p. 0 com/l8cv/BusterX
Abstract
As generative video models become increasingly realistic, detecting AI-generated
videos requires systems that offer both accuracy and interpretability. However,
applying Multimodal Large Language Models (MLLMs) to video forensics is currently limited by outdated datasets, simplistic evaluation protocols, and a reliance
on black-box classification. Experimental results
demonstrate that BusterX outperforms several leading MLLMs (e. 5,
Claude-Sonnet-4. Experimental results
demonstrate that BusterX outperforms several leading MLLMs (e. 5,
Claude-Sonnet-4.Where the paper says it
Where the paper says it
Where the paper says it
Dataset Introduction
The paper introduces GenBuster-200K, a fair dataset of over 200,000 high-quality videos sourced from state-of-the-art generators This dataset is designed to address limitations in prior work by leveraging the latest generators and enforcing rigorous fairness constraints such as balanced gender and ethnicity distributions The construction pipeline involves three stages: Diverse Seed Generation, Description Generation, and Post-filtering
Benchmark Design
GenBuster-Bench is proposed as a diagnostic benchmark that decomposes the video forgery detection challenge along two critical axes: domain shifts and generational shifts This dual-axis design partitions the evaluation into three progressive tracks: In-Domain (Era: 2024), Out-of-Domain (Era: 2025), and In-the-Wild (Era: 2026) The evaluation protocol includes an MLLM-as-a-Judge protocol to assess the quality of the generated forensic explanations
Model Training Strategy
BusterX is developed as an MLLM baseline with Reinforcement Learning (RL) training, formulating detection as a visual reasoning task where the generated reasoning chain serves as the detector itself The RL training utilizes DAPO over GRPO because DAPO effectively prevents entropy collapse inherent to standard GRPO The total reward R is a weighted sum of three components: rformat, raccuracy, and rsoft overlong penalty
Performance Results
Experimental results demonstrate that BusterX outperforms several leading MLLMs (e.g., Qwen3.5, Claude-Sonnet-4.6) in both detection accuracy and rationale quality BusterX achieves the highest overall accuracy among all open-source MLLMs and classical detectors on the In-Domain (ID) track Furthermore, BusterX maintains strong robustness against real-world social media degradations in the Wild track The model’s full responses in Case Study shows that it successfully identifies physical inconsistencies like inconsistent wave progression
Explainability Assessment
BusterX achieves the highest Wild Rationale score as evaluated by the Gemini Judge on correctly predicted fake samples Human evaluation of the rationales generated by BusterX on the Wild track yielded a mean human score of 78.7, which shows a high level of forensic alignment with automated judges The model's ability to pinpoint subtle spatiotemporal inconsistencies is highlighted as a key capability
Future Directions
The current framework is limited to the visual modality, leaving audio manipulation and audiovisual semantic inconsistencies unaddressed Future work will extend toward audio-visual multimodal detection A societal impact note acknowledges the dual-use nature of this research and encourages access controls when deploying forensic tools in adversarial settings The paper also notes a scaling law within the Qwen3.5 family, indicating that larger parameter sizes bring consistent benefits in handling highly diverse, real-world forgeries The study concludes by emphasizing that formulating detection as a step-by-step visual reasoning task is essential The authors also detail the prompt design pipeline used to guide the model's response generation The paper references numerous related works in AI-generated video detection methods and MLLMs The paper concludes by addressing critical vulnerabilities in AI-generated video detection through three contributions The authors also provide detailed experimental details regarding training configurations and evaluation parameters for reproducibility The paper includes a section on related work covering early datasets, detection methods, and MLLMs The paper concludes by providing a list of references spanning various foundational and recent works in the field
--- Page 1 ---
BusterX: MLLM-Powered AI-Generated Video Forgery Detection and Explanation
Forgery Detection and Explanation
Haiquan Wen1∗ Yiwei He1∗ Zhenglin Huang1∗ Tianxiao Li1 Zihan Yu1 Xingru Huang5 Lu Qi4 Baoyuan Wu3 Xiangtai Li2 Guangliang Cheng1†
--- Page 2 ---
Table 1: Comparison with existing datasets. GenBuster outperforms prior works in scale, visual quality and fairness (balanced gender, age, and ethnicity distributions). p. 1 Table 1: Comparison with existing datasets. GenBuster outperforms prior works in scale, visual
quality and fairness (balanced gender, age, and ethnicity distributions).Where the paper says it
--- Page 3 ---
Lack of Forensic Reasoning: Recent methods that adapt MLLMs for detection either use them as auxiliary to assist binary classification, or merely to generate post-hoc explanations Neither approach treats step-by-step physical and spatiotemporal reasoning as the detection mechanism itself
--- Page 4 ---
To address these limitations, we introduce three contributions: 1) GenBuster-200K: A High-Quality and Fair Dataset 2) GenBuster-Bench: A Progressive Diagnostic Benchmark 3) BusterX: A Strong MLLM Baseline with RL Training
--- Page 5 ---
In-Domain (ID) Performance and Bias Mitigation. On the ID track, BusterX achieves the highest overall accuracy among all open-source MLLMs and classical detectors For instance, Qwen3. p. 5 In-Domain (ID) Performance and Bias Mitigation. On the ID track, BusterX achieves the highest
overall accuracy among all open-source MLLMs and classical detectors. For instance, Qwen3. 5-122B exhibits a conservative bias (79. To address these issues, we introduce a comprehensive dataset, benchmark, and baseline model for video forgery detection. First,
we present GenBuster-200K, a fair dataset of over 200,000 high-quality videos
sourced from state-of-the-art generators, featuring diverse real-world scenarios. 7% on Fake). Conversely, Claude-Sonnet-4. Conversely, Claude-Sonnet-4. 6 suffers from an over-forgery bias, which ruins its performance on real
videos (61.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 6 ---
Out-of-Domain (OOD) Generalization. BusterX achieves a remarkable 87. p. 5 The OOD track measures cross-generator generalization exclusively on unseen commercial models. BusterX achieves a remarkable 87. 8% accuracy,
significantly outperforming all classical detectors and MLLMs. In-the-Wild (Wild) Robustness. In-the-Wild (Wild) Robustness. When confronting real-world social media in the Wild track, the
performance of most classical detectors and generic MLLMs severely collapses (e. When confronting real-world social media in the Wild track, the
performance of most classical detectors and generic MLLMs severely collapses (e. , Claude-Haiku4. , Claude-Haiku4. 5 drops to 19.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 7 ---
Is this video REAL or FAKE? The frames show coherent, physically plausible wave motion Conclusion: this looks like a real video. p. 6 :
Is this
video
REAL or
FAKE? The frames show coherent, physically plausible
wave motion.Where the paper says it
--- Page 8 ---
H Additional Examples of BusterX FAKE
--- Page 9 ---
REAL
--- Page 10 ---
GPT-5.2 GPT-5.2 Claude-Sonnet-4. p. 0 Experimental results
demonstrate that BusterX outperforms several leading MLLMs (e. 5,
Claude-Sonnet-4. Experimental results
demonstrate that BusterX outperforms several leading MLLMs (e. 5,
Claude-Sonnet-4. Experimental results
demonstrate that BusterX outperforms several leading MLLMs (e. 5,
Claude-Sonnet-4. Experimental results
demonstrate that BusterX outperforms several leading MLLMs (e. 5,
Claude-Sonnet-4.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 11 ---
BusterX
--- Page 12 ---
[30] Long Ma, Zhiyuan Yan, Qinglang Guo, Yong Liao, Haiyang Yu, and Pengyuan Zhou. Detecting aigenerated video via frame consistency
--- Page 13 ---
B Prompt Generation Pipeline in GenBuster-200K
Stage-1: Diverse Seed Generation. This stage is explicitly engineered for Fairness Stage-2: Description Generation. p. 0 Specifically, we
identify these critical limitations in the current research landscape: 1) Outdated and Biased Data:
Existing video datasets [5, 30, 9, 32, 18, 19] frequently rely on early-stage generators, producing easily
recognizable artifacts. Furthermore, they often overlook demographic fairness, failing to account
for gender and racial diversity. In the video domain, MM-Det [44] leverages the strong
visual-linguistic representations of MLLMs by generating a Multi-Modal Forgery Representation,
which enhances its ability to detect unseen manipulated content. Reinforcement Learning (RL)
has demonstrated its effectiveness in enhancing LLM reasoning capabilities [1, 49] following the
Supervised Fine-Tuning (SFT) stage [29].Where the paper says it
Where the paper says it
--- Page 14 ---
C DAPO
We utilize DAPO [57] as our reinforcement learning strategy JDAPO(θ) = E(q,a)∼D,min ri,t(θ)Aˆi,t,clipri,t(θ), 1 − εlow, 1 + εhigh Aˆi,t is the advantage
--- Page 15 ---
F MLLM-as-a-judge
We employ Gemini 3 Flash Preview as Judge The final explainability score is derived by averaging these dimensions and strictly normalizing the result to the [0, 1] interval
--- Page 16 ---
G Complete Responses
GPT-5.2 GPT-5.2 Claude-Sonnet-4.6 Claude-Sonnet-4.6 BusterX BusterX
--- Page 17 ---
BusterX BusterX
--- Page 18 ---
H Additional Examples of BusterX FAKE
--- Page 19 ---
REAL
The paper introduces GenBuster-200K, a fair dataset of over 200,00K high-quality videos sourced from state-of-the-art generators This dataset is designed to address limitations in prior work by leveraging the latest generators and enforcing rigorous fairness constraints such as balanced gender and ethnicity distributions GenBuster-Bench is proposed as a diagnostic benchmark that decomposes the video forgery detection challenge along two critical axes: domain shifts and generational shifts BusterX is developed as an MLLM baseline with RL training, formulating detection as a visual reasoning task where the generated reasoning chain serves as the detector itself
Pruning redundant reasoning tokens
The paper introduces GenBuster-200K, a fair dataset of over 200,0K high-quality videos sourced from state-of-the-art generators This dataset is designed to address limitations in prior work by leveraging the latest generators and enforcing rigorous fairness constraints such as balanced gender and ethnicity distributions GenBuster-Bench is proposed as a diagnostic benchmark that decomposes the video forgery detection challenge along two critical axes: domain shifts and generational shifts BusterX is developed as an MLLM baseline with RL training, formulating detection as a visual reasoning task where the generated reasoning chain serves as the detector itself
How the training signal is built
The RL training utilizes DAPO over GRPO because DAPO effectively prevents entropy collapse inherent to standard GRPO The total reward R is a weighted sum of three components: rformat, raccuracy, and rsoft overlong penalty The reward function components are defined as follows: Format (rformat) receives a reward of 0 if the model adheres to the specified format or a penalty of −1 otherwise Accuracy (raccuracy) receives a reward of 1 if formatted and correct, otherwise it receives 0 Soft Overlong Penalty penalizes excessively long responses to prevent computational waste and mitigate repetitive token generation
Where the gains come from
BusterX achieves the highest overall accuracy among all open-source MLLMs and classical detectors on the In-Domain (ID) track BusterX demonstrates that domain-specific MLLMs can maintain strong robustness against real-world social media degradations in the Wild track The performance gap between generic models and BusterX illustrates that RL-driven visual reasoning is crucial for bridging the gap to real-world applications BusterX achieves the highest Wild Rationale score as evaluated by the Gemini Judge on correctly predicted fake samples This confirms that BusterX not only identifies forgeries accurately but also provides high-quality explanations The paper also notes a scaling law within the Qwen3.5 family, indicating that larger parameter sizes bring consistent benefits in handling highly diverse, real-world forgeries The study concludes by emphasizing that formulating detection as a step-by-step visual reasoning task is essential The authors also provide detailed experimental details regarding training configurations and evaluation parameters for reproducibility The paper includes a section on related work covering early datasets, detection methods, and MLLMs The paper concludes by providing a list of references spanning various foundational and recent works in the field
--- Page 19 ---
REAL</ref:2505.
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
-
Enhanced detection accuracy via BusterX by formulating detection as a visual reasoning task where
the generated reasoning chain serves as detector itself.
This allows models to uncover physical inconsistencies without expensive text annotations. -
Improved generalization across different generative models using GenBuster-Bench, which decomposes the challenge along two critical axes: domain shifts and generational shifts (In-Domain, Out-of-Domain, and In-the-Wild tracks).
-
Superior forensic explanation quality through the MLLMas Judge protocol, which assesses
the forensic grounding and reasoning depth of the MLLMs
on samples from the Wild track using a structured 1 to 5 scale. -
Mitigated bias in detection models by training BusterX with Reinforcement Learning (RL) and employing reward functions that enforce format and accuracy, ensuring
a balanced performance, demonstrating that our training paradigm successfully aligns the model’s forensic perception.
-
Increased robustness against real-world degradations by testing models across various conditions defined in Table 4, showing consistent performance even under
JPEG 80
orJPEG 70
degradation.
Abstract
As generative video models become increasingly realistic, detecting AI-generated videos requires systems that offer both accuracy and interpretability. However, applying Multimodal Large Language Models (MLLMs) to video forensics is currently limited by outdated datasets, simplistic evaluation protocols, and a reliance on black-box classification. To address these issues, we introduce a comprehensive dataset, benchmark, and baseline model for video forgery detection. First, we present GenBuster-200K, a fair dataset of over 200,000 high-quality videos sourced from state-of-the-art generators, featuring diverse real-world scenarios. Second, we propose GenBuster-Bench, a diagnostic benchmark spanning three progressive tracks (In-Domain, Out-of-Domain, and In-the-Wild) to evaluate models across domain shifts and generational shifts. Finally, we develop BusterX, a simple MLLM baseline with Reinforcement Learning (RL) training. Instead of direct binary classification, BusterX formulates detection as a visual reasoning task, where the generated reasoning chain serves as detector itself. Experimental results demonstrate that BusterX outperforms several leading MLLMs in both detection accuracy and rationale quality.
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs