FakeVLM-R1: Internalizing Physical Laws via CoT for Synthetic Image Detection

Source: arXiv:2605.30062 · Published 2026-05-28 · By Leqi Zhu, Junyan Ye, Kaiqing Lin, Zhiyuan Yan, Conghui He, Weijia Li

TL;DR

The paper addresses the critical challenge of synthetic image detection amid rapidly improving generative AI capabilities that produce highly realistic fake images. Existing detection methods mainly rely on imitation learning from large forged datasets, lacking true causal reasoning and suffering from explanatory hallucinations and over-rejection of real images. To overcome these limitations, the authors propose FakeVLM-R1, a novel framework that incorporates reinforcement learning via Group Relative Policy Optimization (GRPO) and a bidirectional Critical Thinking Chain-of-Thought (CoT) mechanism. This enables the model to simultaneously generate forgery hypotheses and authenticity counter-proofs grounded in physical commonsense during inference, dramatically improving logical consistency and reducing false positives.

Complementing this, they introduce FakeClue++, a curated 50K-sample dataset enriched with fine-grained artifact annotations and detailed physical law-based authenticity explanations for both synthetic and real images. Experiments show FakeVLM-R1 achieves state-of-the-art detection accuracy, surpasses previous explainable LMM baselines, and improves generalization to unseen domains and robustness against perturbations. The paradigm shift from passive pattern recognition to active dialectical reasoning enables forensic-level explanations with practical low false positive risk on real-world images.

Key findings

• FakeVLM-R1 reduces hallucination and over-rejection bias on real images compared to FakeVLM and Gemini-2.5 (Fig. 4 & 5).
• FakeClue++ dataset contains 50K samples (roughly balanced fake/real) with expert-verified artifact and physical law annotations, enabling higher data efficiency than the original 100K FakeClue dataset.
• The bidirectional dialectical CoT with GRPO reinforcement learning yields better logical consistency and self-consistency of explanations, measured by logical consistency scores from an external critic SophiaVL-R1.
• FakeVLM-R1 attains state-of-the-art detection accuracy and explanation quality across multiple benchmarks, surpassing prior frameworks that rely solely on supervised fine-tuning or imitation learning.
• Incorporating authenticity counter-proofs based on physical laws (e.g., lighting, perspective) provides robust anchors that reduce false alarms in natural noisy real images.
• GRPO reinforcement learning optimizes multiple reward components (correctness, format, structural dialectic conformity, logical consistency, and reasoning length) to enhance bidirectional reasoning depth.
• FakeVLM-R1 achieves superior generalization on out-of-domain data despite only using half the volume of training data compared to the original FakeClue.
• The critical thinking paradigm successfully enforces self-consistent forensic reasoning via explicit <think> tags pairing each artifact clue with both fake and real counterarguments.

Threat model

Adversary: advanced image forgers using state-of-the-art generative AI to produce highly realistic fake images. Defender’s goal is to detect such synthetic images using multimodal reasoning. The adversary cannot manipulate the detection pipeline directly or produce invulnerable synthetic images with full physical consistency. The defender assumes access to both forged and authentic images with rich annotations and relies on internalizing physical laws for detection.

Methodology — deep read

The authors present a two-stage training framework building upon large multimodal models (LMMs) to enhance synthetic image detection through logical reasoning grounded in physical laws.

1. Threat Model & Assumptions: The adversary is an advanced image forger using state-of-the-art generative models producing realistic fakes. The defender has access to both forged and authentic images but assumes that purely perceptual cues are insufficient, requiring internalization of physical commonsense. The adversary cannot manipulate the detection framework or provide adversarial examples beyond natural image variants.

2. Data: FakeClue++ dataset with 50,000 samples, balanced between fakes and real images, drawn from multiple sources including the previous FakeClue dataset, SynthScars, OpenImages, and internet photography. Data cleaned and verified by human experts. Labeled with fine-grained regions of artifacts and structured explanatory annotations highlighting physical inconsistencies (lighting, shadows, anatomy). Dataset split into training for Supervised Fine-Tuning (SFT) and a 4,000-image benchmark test set for robustness/generalization evaluation.

3. Architecture and Algorithm: Base architecture is a vision-language model (e.g., Qwen-VL or LLaVA) with a standard vision encoder feeding embeddings into a large language model (LLM) decoder. Stage 1 uses SFT with autoregressive loss on the reformulated diverse explanation prompts. Stage 2 replaces passive imitation with reinforcement learning using Group Relative Policy Optimization (GRPO), which trains the model to produce dialectical bidirectional reasoning chains-of-thought (CoT). The CoT requires the model to generate pairs of reasoning per artifact: a forgery hypothesis ([Why fake]) and a simultaneous authenticity counter-proof ([If real]), or vice versa. This forces rigorous logical self-consistency before concluding fake/real verdicts.

4. Training: Stage 1 Cold Start involves supervised fine-tuning for 5-10 epochs (details on batch size, optimizer unreported) to establish artifact recognition and explanation generation capability. Stage 2 applies GRPO to optimize multiple reward signals (correctness, formatting, dialectical structure, logical consistency evaluated by external critic SophiaVL-R1, and reasoning chain length). GRPO avoids critic network by intra-group advantage calculation using sampled reasoning paths and their reward statistics. The model is updated to maximize expected group-relative advantage.

5. Evaluation: Metrics include detection accuracy, false positive rate on real images, and logical consistency scores from critic models. Experiments compare FakeVLM-R1 versus FakeVLM (the preliminary version without reinforcement) and other LMM baselines like Gemini-2.5 across multiple mainstream benchmarks and on the held-out FakeClue++ test set to assess generalization and robustness to perturbations. Ablations isolate the effects of bidirectional dialectic reasoning and physical law annotations.

6. Reproducibility: The authors mention leveraging open-source frameworks (vLLM and SGLang) for inference but do not clarify if code or pretrained weights will be released. Dataset sources are partially public (OpenImages, SynthScars) but include proprietary expert annotations. Method details on hyperparameters are limited.

Concrete Example: For an input image suspected fake, the model generates a <think> tag containing paired statements for each artifact clue (e.g., unusual skin texture). It states why the anomaly suggests forgery, then immediately counters with physical law reasons it could be authentic (e.g., lighting or anatomy plausibility). After iterating over clues, it integrates this self-consistency check to output <answer> Fake or Real. Training reinforcement maximizes rewards for reasoning chains that fulfill this bidirectional dialectical format, hold logical consistency, and yield correct detection results.

Technical innovations

• Integration of Group Relative Policy Optimization (GRPO) reinforcement learning into LMM fine-tuning to enable bidirectional dialectical reasoning.
• Novel Critical Thinking Chain-of-Thought (CoT) framework requiring paired forgery hypotheses and authenticity counter-proofs for each artifact clue.
• Creation of FakeClue++, a synthetic detection dataset combining expert-labeled fine-grained artifact annotations with detailed physical law-based authenticity explanations for real and fake images.
• Composite multi-dimensional reward design enforcing correctness, dialectical structure, logical consistency (via external VLM critic), and efficient reasoning length during RL training.

Datasets

• FakeClue++ — 50,000 samples — curated from FakeClue, SynthScars, OpenImages, social media, with expert-verified annotations

Baselines vs proposed

• FakeVLM (Preliminary): detection accuracy lower than FakeVLM-R1 by significant margin; more hallucinations and higher false positive rate on real images.
• Gemini-2.5: lower accuracy and logical explanation consistency compared to FakeVLM-R1 (Fig. 4, Fig. 5).
• Supervised Fine-Tuning only (no GRPO): inferior logical consistency and higher hallucination risk versus SFT + GRPO (ablation studies).

Figures from the paper

Figures are reproduced from the source paper for academic discussion. Original copyright: the paper authors. See arXiv:2605.30062.

Fig 1: Comparison of synthetic detection paradigms. (a) Tra-

Fig 2: Overview of the FakeClue++ dataset. Diverging from traditional datasets focused on artifact annotation, FakeClue++

Fig 5: Comparison of reasoning processes on fake Samples among FakeVLM-R1, FakeVLM and Gemini-2.5.

Fig 6: Comparison of reasoning processes on real Samples among FakeVLM-R1, FakeVLM and Gemini-2.5.

Fig 4: Fine-grained comparison in FakeClue++. Performance

Fig 6 (page 10).

Fig 7 (page 10).

Fig 8 (page 10).

Limitations

• Limited details disclosed on training hyperparameters, model size, and hardware used, impeding reproducibility.
• No adversarial robustness evaluation against intentional forger attacks or adaptive adversaries.
• Physical priors are limited to certain annotated attributes; complex or subtle real-world scene variations may still cause misclassifications.
• Dataset sources partly drawn from proprietary curation; accessibility of FakeClue++ for external researchers unclear.
• Evaluation mostly focuses on existing benchmarks and held-out subsets; broader real-world deployment scenarios not fully explored.

Open questions / follow-ons

• How resilient is FakeVLM-R1 to adaptive adversarial manipulations that attempt to mimic physical laws in fakes?
• Can the bidirectional CoT approach be extended effectively to video or multi-frame synthetic content detection?
• What is the tradeoff between reasoning chain length and inference latency in real-time deployment scenarios?
• How well does the framework generalize to completely novel synthetic generation architectures unseen during training?

Why it matters for bot defense

For practitioners designing bot defense and CAPTCHA systems, this work demonstrates a significant advancement in interpretability and robustness of synthetic image detection. Integrating physical commonsense reasoning and bidirectional dialectical logic into large multimodal models can greatly reduce false positives against real images—a common challenge in bot detection when distinguishing human-generated visual input from AI-generated fakes. The FakeVLM-R1 paradigm highlights that relying solely on visual artifact pattern recognition is insufficient for high-fidelity forgery detection in complex real-world scenarios. Bot defense engineers might consider adopting or adapting similar critical thinking and reinforcement learning approaches to enhance explainability and confidence in forgery judgments. Additionally, the dataset construction approach emphasizes the value of combining synthetic artifact supervision with explicit physical law annotations to build models that internalize authenticity anchors rather than just memorizing forgery fingerprints. However, practitioners should also recognize the current limitations such as the lack of adversarial robustness testing and partial dataset accessibility when evaluating integration into production bot defense pipelines.

Cite

@article{arxiv2605_30062,
  title={ FakeVLM-R1: Internalizing Physical Laws via CoT for Synthetic Image Detection },
  author={ Leqi Zhu and Junyan Ye and Kaiqing Lin and Zhiyuan Yan and Conghui He and Weijia Li },
  journal={arXiv preprint arXiv:2605.30062},
  year={ 2026 },
  url={https://arxiv.org/abs/2605.30062}
}

Read the full paper

• arXiv:2605.30062 (abstract)
• arXiv:2605.30062 (PDF)