Veritas++: Value-aware On-Policy Distillation for Perception-Enhanced AIGI Detection

Source: arXiv:2607.27113 · Published 2026-07-29 · By Hao Tan, Jun Lan, Zichang Tan, Ajian Liu, Zijian Yu, Chuanbiao Song et al.

TL;DR

This paper addresses the challenge of detecting AI-generated images (AIGI) by improving multimodal large language models (MLLMs) with stronger visual perception capabilities. Current MLLM-based detectors focus mainly on organizing visual evidence in their reasoning but suffer from perception bottlenecks that limit capturing fine-grained details, semantic anomalies, and pixel-level differences critical for robust detection. Veritas++ introduces a perception-enhanced reasoning framework that grounds detection on these three core perceptual abilities through Perception-oriented Learning (PoRL), which replaces open-ended supervision with verifiable rewards. Additionally, it incorporates Value-aware On-Policy Distillation (VaOPD), an adaptive self-distillation approach that prioritizes high-value distillation signals to internalize perception-aware reasoning efficiently. Experimental results across multiple benchmarks demonstrate that Veritas++ significantly reduces perception gaps, raising semantic anomaly detection accuracy from around 62.7% to 97.4%, and boosts detection performance while enabling continual capability evolution.

Key findings

• Veritas++ (Perception-oriented Learning) achieves 97.4% average accuracy on semantic anomaly and pixel-level perception benchmarks (AbHuman, SDG, MagicData, B-Free), improving over baseline VERITAS (64.6% avg) and other methods such as FakeVLM (72.4%) and DeepVRM (86.9%).
• Detection ACC improves to 86.9% with perception-optimized Veritas++, compared to 71.5% for VERITAS and 70.3–76.7% for other baselines.
• PoRL uses verifiable rewards for fine-grained visual details, semantic anomalies, and pixel discrepancies to explicitly reinforce foundational perceptual capacities.
• VaOPD reweights distillation loss at trajectory, token, and directional levels based on privileged teacher signals, emphasizing erroneous rollouts and tokens with high teacher–student divergence for more effective reasoning internalization.
• Filtering for high-quality human-annotated explanations during cold start reduces hallucination and forms a reliable perceptual and reasoning prior, enabling better downstream learning.
• VaOPD enables efficient model capability evolution to novel AIGI generators and scenarios without sacrificing previously learned detection performance.
• Pixel-level perception is improved by learning to localize subtle inpainting artifacts using IoU and area-penalty rewards, overcoming challenges in describing low-level signal anomalies via language reasoning.
• Evaluations include diverse datasets: AbHuman (anatomical anomalies), MagicData (semantic anomalies), SDG and B-Free (pixel-level differences), demonstrating broad generalization gains.

Threat model

The adversary is an image generator aiming to produce photorealistic yet synthetic images that evade detection by masking visual and semantic anomalies. The adversary can generate diverse and high-fidelity forgeries but cannot perfectly eliminate all fine-grained artifacts or pixel-level inconsistencies. The defender uses a multimodal large language model-based detector that must perceive subtle visual cues and anomalies to reliably classify images as real or fake, without the use of external privileged evidence at inference time.

Methodology — deep read

1. Threat Model & Assumptions: The framework assumes an adversary generating synthetic images with subtle semantic or pixel-level anomalies aiming to evade detection. The adversary can produce high-quality forgeries but cannot completely remove all detectable artifacts or anomalies. The detector must identify fine-grained visual defects and semantic violations using perceptually grounded reasoning without external privileged modules at test time.

2. Data: The approach aggregates a curated dataset (~10K samples) of images with manual artifact annotations from SynthScars, AbHuman, SynthArtifact, X-AIGD, and HydraFake, enriched with real images from Flickr and VisualGenome for a high-quality cold start. The annotations include short explanations, forgery masks, and detailed chain-of-thought (CoT) trajectories, all rewritten into pattern-aware reasoning style guided by Gemini-2.5-Pro and filtered by GPT-5.1 to remove hallucinations.

3. Architecture / Algorithm: Starting from a base multimodal large language model (MLLM), Veritas++ trains in three stages. Stage 1: Supervised fine-tuning on the high-quality cold-start data to establish a low-hallucination reasoning prior.

Stage 2: Perception-oriented Learning (PoRL) involves targeted tasks to enhance three perception abilities: (a) Fine-grained visual perception using questions on localized image regions from ZwZ and Vero datasets, rewarded by exact or LLM-judged answer matches. (b) Semantic anomaly perception via region-level normal/abnormal body part classification (AbHuman) and image-level anomaly categorization (MagicData), rewarded by structured label matching. (c) Pixel discrepancy perception using inpainted/localized modifications from Next-IMDL dataset, rewarded by Intersection-over-Union (IoU) with area penalties and overall image authenticity correctness.

Rewards are combined with length penalties and optimized with a policy gradient algorithm (GSPO).

Stage 3: Value-aware On-Policy Distillation (VaOPD) enables a privileged self-teacher providing dense token-level distributions conditioned on privileged information (including training labels). The student samples reasoning trajectories under standard queries. VaOPD models the distillation loss at three levels: trajectory (upweighting incorrect rollouts), token (weighting tokens with large KL divergence between teacher and student), and adaptive direction (using entropy difference and teacher advantage to modulate distillation direction). The goal is to internalize perception-enhanced reasoning efficiently.

4. Training Regime: Cold start fine-tuning on ~10K high-quality annotations optimizes cross-entropy token likelihood. For PoRL, policy optimization with multiple sampled responses per query uses reward signals as above. VaOPD applies self-distillation with an exponential moving average (EMA) teacher. Specific hyperparameters (batch size, epochs, seed, hardware) are not explicitly detailed.

5. Evaluation Protocol: Perception abilities are evaluated on four specialized datasets: AbHuman (anatomical anomaly classification), MagicData (structural anomaly detection), SDG, and B-Free (pixel-level discrepancy). Detection accuracy is measured by final real/fake classification accuracy. Comparisons include several baselines: Qwen3-VL-8B, UniGenDet, FakeVLM, DeepVRM, and VERITAS. Ablations analyze intermediate policies (SFT only) and progressive additions (perceptual optimization, VaOPD). Held-out real-world and emerging generative model test sets are included for generalization assessment.

6. Reproducibility: The authors provide code and checkpoints publicly at https://github.com/EricTan7/VeritasPP. Dataset provenance is detailed but some sets (e.g., HydraFake MiPO) may be restricted. Model details like base MLLM and training seeds are partially described, with some reliance on proprietary LLMs (Gemini, GPT-5.1) for annotation rewriting and filtering. Exact replication may require access to these tools.

Example End-to-End: Given an input image suspected of AIGI, Veritas++ first generates structured reasoning that includes explicit attention to fine-grained evidence (e.g., a distorted limb identified via bounding box), semantic anomaly classification, and pixel-level discrepancy localization. Perception-oriented Learning has enabled the model to verify these anomalies via reward-optimized policies. Then, VaOPD distills privileged teacher feedback emphasizing correction of perceptual errors back into the student model's reasoning steps, resulting in a robust and interpretable authenticity verdict.

Technical innovations

• Introduction of Perception-oriented Learning (PoRL) applying verifiable reward signals instead of open-ended description supervision to explicitly enhance fine-grained visual detail, semantic anomaly, and pixel-level discrepancy perception within MLLMs.
• Proposal of Value-aware On-Policy Distillation (VaOPD), an adaptive self-distillation method that weights distillation loss by trajectory correctness, token-level teacher-student divergences, and adaptive directional signals to better integrate perception-enhanced reasoning.
• Design of a low-hallucination high-quality cold start dataset construction pipeline by retaining only human-annotated artifact explanations and rewriting them into pattern-aware reasoning chains with quality filtering, ensuring reliable perception and reasoning foundation.
• Application of multi-level adaptive reweighting in distillation (trajectory, token, direction) to enable efficient internalization and continual capability evolution for detection of novel AI-generated image types.

Datasets

• AbHuman — size unspecified — anatomical anomaly detection with region-level normal/abnormal labels
• MagicData — size unspecified — image-level semantic anomaly detection with multi-level anomaly categories
• SDG — size unspecified — pixel-level discrepancy detection
• B-Free — size unspecified — pixel-level discrepancy detection using authentic and locally inpainted image pairs
• SynthScars, SynthArtifact, X-AIGD, HydraFake MiPO portion (~10K samples combined) — high-quality human-annotated explanations/forgery masks
• ZwZ and Vero datasets — size unspecified — for fine-grained visual perception via localized questions
• Next-IMDL — size unspecified — dataset of real and inpainted image pairs for pixel discrepancy perception

Baselines vs proposed

• Qwen3-VL-8B: avg perception task accuracy = 53.1% vs Veritas++ (Perception) = 96.6%
• UniGenDet: avg perception task accuracy = 52.0% vs Veritas++ (Perception) = 96.6%
• FakeVLM: avg perception task accuracy = 72.4% vs Veritas++ (Perception) = 96.6%
• DeepVRM: avg perception task accuracy = 86.9% vs Veritas++ (Perception) = 96.6%
• VERITAS baseline: avg perception task accuracy = 64.6% vs Veritas++ (SFT only) = 76.4% vs Veritas++ (Perception) = 96.6%
• Detection accuracy: Qwen3-VL-8B = 70.3%, UniGenDet = 70.5%, FakeVLM = 71.2%, DeepVRM = 76.7%, VERITAS = 71.5%, Veritas++ (SFT) = 82.2%, Veritas++ (Perception) = 86.9%

Figures from the paper

Figures are reproduced from the source paper for academic discussion. Original copyright: the paper authors. See arXiv:2607.27113.

Fig 1: Paradigms of MLLM-based AIGI detection. (a) Explanation-supervised methods learn from artifact rationales.

Fig 2: Examples of perception challenges in AIGI detec-

Fig 3: Overview of VERITAS++. Stage 1: High-quality Cold Start establishes a reliable, low-hallucination reasoning prior.

Fig 4 (page 2).

Fig 5 (page 2).

Fig 6 (page 2).

Fig 7 (page 2).

Fig 4: Empirical motivation for VaOPD. Token-wise teacher

Limitations

• The cold-start dataset is limited (~10K samples) and relies heavily on diverse but manually annotated labels, potentially constraining scalability and domain coverage.
• Perception reward functions depend on exact or LLM judgment-based matching that might fail on subtle or ambiguous cases, risking noisy supervision.
• Some core components (annotation rewriting by Gemini-2.5-Pro and filtering by GPT-5.1) rely on proprietary models not publicly available, hindering full reproducibility.
• Evaluation focuses on standard and emerging detection benchmarks, but lacks extensive adversarial robustness tests against adaptive synthetic image generators.
• Computation cost and latency of VaOPD and multi-stage training pipeline are not reported, which may impact deployment.
• The framework depends on a privileged self-teacher during distillation, whose quality and biases strongly affect final performance.

Open questions / follow-ons

• Can Veritas++ perception and reasoning improvements be extended to video-based generative content for deepfake detection?
• How robust is the method against adaptive adversaries specifically targeting the perceptual abilities reinforced by PoRL and VaOPD?
• Can the framework be generalized to other modalities (e.g., audio deepfakes) where fine-grained anomaly perception is critical?
• What are the trade-offs between model size, inference latency, and detection accuracy in large-scale deployment of Veritas++?

Why it matters for bot defense

For bot-defense and CAPTCHA practitioners, Veritas++ offers a meaningful advancement in generating interpretable, perception-aware models for distinguishing synthetic visual content. Since synthetic images may be used for sophisticated bot profiles or phishing attacks, improving detectors' capability to identify fine-grained anomalies and semantic inconsistencies strengthens overall defense. The PoRL methodology highlights the importance of grounding model supervision on verifiable rewards that specifically target perceptual weaknesses rather than relying on ambiguous or open-ended explanations, which is a useful design insight for future detection systems. Furthermore, VaOPD shows a promising approach to efficiently internalizing expert feedback in a continuous learning setup, enabling detectors to evolve alongside emerging generative models. Nonetheless, deployment teams should weigh computational demands and consider adversarial robustness for real-world scenarios.

Cite

@article{arxiv2607_27113,
  title={ Veritas++: Value-aware On-Policy Distillation for Perception-Enhanced AIGI Detection },
  author={ Hao Tan and Jun Lan and Zichang Tan and Ajian Liu and Zijian Yu and Chuanbiao Song and Huijia Zhu and Weiqiang Wang and Jun Wan and Zhen Lei },
  journal={arXiv preprint arXiv:2607.27113},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.27113}
}

Read the full paper

• arXiv:2607.27113 (abstract)
• arXiv:2607.27113 (PDF)