acceptodds
Withdrawn from ICLR 2027

ShopEye: Bridging Agentic Data Construction and Active Visual Verification for Fine-Grained E-Commerce AIGC Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The rapid advancement of AIGC technologies provides a convenient and effective means for product display, driving an exponential growth of AI-generated content on e-commerce platforms. While substantially boosting page views and transaction conversion rates, it introduces uncontrollable generative artifacts, such as distorted product features, severely impairing the user experience. However, existing benchmarks are predominantly confined to discerning whether an image is AI-generated, rather than conducting fine-grained quality assessment on such generated content. To bridge this gap, we present ShopEye, a full-lifecycle agentic framework for fine-grained AIGC defect diagnosis in e-commerce. ShopEye integrates two core components: ShopEye-Agent for automated data production, and ShopEye-Trainer for agentic model training. Specifically, ShopEye-Agent employs an automated multi-agent pipeline to generate and curate diagnostic data. This produces ShopEye-Bench, the first benchmark tailored for low-quality e-commerce AIGC, comprising 5K training and 1K test samples. It covers prevalent defect types in e-commerce scenarios, including product inconsistency, human deformity, unrealistic scenes, and text distortion. Nonetheless, detecting these defects typically requires fine-grained visual understanding. To enhance the model's diagnostic capability, ShopEye-Trainer incorporates diverse visual tools for perception augmentation, and optimizes the diagnostic policy via reinforcement learning. Specifically, it couples Local-Global Dual-View Supervised Finetuning to model fine-grained visual details, and Evidence-Weighted Policy Optimization to incentivize active visual tool invocation for evidence gathering. By combining ShopEye-Bench with our training paradigm, we develop ShopEye-4B and ShopEye-8B, two specialized multimodal agentic models at different scales. Extensive experiments demonstrate that both models consistently outperform strong baselines, validating the effectiveness and scalability of our training framework with only a modest amount of task-specific training data.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Agents, Fine-Grained Visual Understanding, Tool-Augmented Reasoning, Datasets and Benchmarks

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_q9UmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ShopEye is a model for detecting defects in AIGC-generated images in e-commerce scenarios. The paper constructs valuable datasets across dimensions such as Product, Human, Scene, and Text, and proposes an agent system, forming a relatively complete engineering system. However, the two core techniques proposed in ShopEye-Trainer both seem to be simple improvements over existing methods. The paper does not propose a new architecture, nor does it fundamentally change the RL optimization paradigm. ShopEye conducts targeted training and evaluation on e-commerce data (especially VTON data). Although it achieves better results compared to other datasets, this only demonstrates the effect of the data rather than the effect of the method.

Critical strengths
  1. Detecting defects in AIGC-generated images is a valuable problem in e-commerce scenarios. ShopEye conducts targeted training and evaluation on e-commerce data (especially VTON data) and achieves good results compared to other methods. ShopEye is a useful application tool.
Critical weaknesses
  1. The two core techniques proposed in ShopEye-Trainer both appear to be simple improvements over existing methods: LGDV-SFT is essentially a combination of multi-turn tool-calling supervision and local visual supervision; Evidence-Weighted Policy Optimization essentially multiplies the existing DAPO advantage estimates for tool-calling tokens by an external verifier score. The paper does not propose a new architecture, nor does it fundamentally change the RL optimization paradigm.
  2. ShopEye conducts targeted training and evaluation on e-commerce data (especially VTON data). Although it achieves better results compared to other datasets, this only demonstrates the effect of the data rather than necessarily the effect of the method. The authors could supplement the experiments with training results of a few other methods on the same dataset.
  3. The data distribution is relatively narrow, and the generalization ability has not been sufficiently validated. Although the paper claims to address AIGC-generated image anomaly detection in e-commerce scenarios, the data is actually dominated by VTON data.
  4. The paper lacks a quantitative evaluation of true "fine-grained localization capability." Although the paper repeatedly emphasizes Fine-Grained, Evidence-Grounded, and Localized Diagnosis, it does not actually provide evaluations of these capabilities. This means that the model may receive a correct score simply by identifying an anomaly in the Human dimension of an image, even if it localizes the wrong region or provides an incorrect visual explanation.
Questions for authors
  1. Supplement experiments on SFT training of other methods using the ShopEye e-commerce dataset. Only a small number of models need to be evaluated (e.g., Qwen3-VL-8B with SFT), so no large-scale experiments are needed.
  2. The data distribution is relatively narrow, and the generalization ability has not been sufficiently validated. Could the authors evaluate the method on more e-commerce datasets? If time is insufficient, this experiment can be omitted.
  3. The paper lacks a quantitative evaluation of true "fine-grained localization capability." Although the paper repeatedly emphasizes Fine-Grained, Evidence-Grounded, and Localized Diagnosis, it does not actually provide evaluations of these capabilities. This evaluation needs to be added.

These experiments would help improve the final score.

Reviewer_YxGVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ShopEye, a framework for fine-grained AIGC defect diagnosis in e-commerce. It consists of ShopEye-Agent, a multi-agent pipeline for automated diagnostic data construction; ShopEye-Bench, a benchmark with 5K training and 1K test samples covering four defect dimensions; and ShopEye-Trainer, which combines LGDV-SFT and EW Policy Optimization to improve active visual verification. Experiments demonstrate that the resulting ShopEye-4B and ShopEye-8B models outperform strong baselines on the proposed benchmark.

Critical strengths
  1. The paper addresses a practically meaningful problem. E-commerce imagery must preserve product attributes while maintaining plausible human anatomy, scenes, and text. The proposed four-dimensional task formulation captures localized defects that may be overlooked by overall image quality assessment. This setting offers a valuable application for studying fine-grained multimodal understanding.
  2. Comprehensive Framework. The paper presents a coherent pipeline connecting automated data construction, benchmark development, and model training.
  3. Strong Experimental Performance. ShopEye-4B and ShopEye-8B achieve Overall scores of 76.34 and 76.83, substantially outperforming their respective base models and competitive baselines.
  4. Clear Writing and Presentation. The paper is generally well written and easy to follow. The motivation, framework design, and training objectives are clearly presented.
Critical weaknesses
  1. Limited Fine-Grained Evaluation. Although the paper emphasizes identifying whether, where, and why defects occur, the primary evaluation focuses on four-dimensional binary classification using Precision, Recall, and F1. Quantitative evaluations of localization accuracy, evidence correctness, and explanation faithfulness are missing, leaving the central claims about evidence-grounded diagnosis insufficiently validated.
  2. Limited Methodological Novelty and Gains. Combining LGDV and EW improves Overall performance by only 0.08 percentage points over EW alone (76.83 vs. 76.75), while EW improves over the correctness-gated baseline by 0.64 points. The independent contributions and robustness of these improvements remain unclear.
  3. Benchmark Reliability and Generalization. ShopEye-Bench primarily relies on an internal apparel virtual try-on pipeline, limiting its demonstrated applicability to broader e-commerce scenarios. And detailed inter-annotator agreement results and test-set adjudication procedures are not reported.
  4. Experimental Inconsistencies. Several discrepancies affect the clarity and reproducibility of the experiments. Section 6.2 reports an Overall score of 77.54 for ShopEye-8B, whereas Table 1 reports 76.83. Section 6.1 allows ShopEye up to 6 tool calls during inference, while Appendix D.1.1 limits baselines to 4 tool calls and 6 model actions. Since model actions are not equivalent to tool calls, this raises concerns about whether the comparisons use consistent interaction budgets. Additionally, the reported LoRA ranks differ between the main implementation details and appendix analyses (64 vs. 128).
Questions for authors
  1. Please provide quantitative evaluations of defect localization, visual evidence correctness, and explanation faithfulness to better support the claims of fine-grained, evidence-grounded diagnosis.
  2. Given the limited improvement of LGDV+EW over EW alone, could the authors clarify the methodological novelty and provide more experiments or statistical analysis to establish the robustness of the reported gains?
  3. Please provide more details on ShopEye-Bench annotation reliability, including inter-annotator agreement, test-set adjudication.
  4. Could the authors clarify the inconsistencies in the reported performance, tool-call budgets, and LoRA configurations? Do these discrepancies affect the main experimental conclusions?
Reviewer_LMVQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses fine-grained AIGC diagnosis in e-commerce and proposes ShopEye, an agentic framework integrating automated data construction and active visual verification. Specifically, ShopEye-Agent employs a multi-agent pipeline to construct ShopEye-Bench, while ShopEye-Trainer combines Local-Global Dual-View SFT with Evidence-Weighted Policy Optimization to improve visual inspection and defect reasoning. Experiments on ShopEye-Bench report substantial improvements over general-purpose and agentic multimodal baselines.

Critical strengths
  1. The motivation is good. The paper proposes an important problem in e-commerce AIGC quality assessment. ShopEye-Bench extends conventional image-level detection toward multi-dimensional defect diagnosis with localized evidence and reasoning trajectories.

  2. The proposed agent system is comprehensive, and performance is good. The proposed framework integrates automated data construction, visual tool use, and reinforcement learning into a unified pipeline. The reported results show substantial improvements over various multimodal baselines, demonstrating the potential value of specialized training for fine-grained visual inspection.

Critical weaknesses
  1. Limited methodological novelty. The overall framework appears largely engineering-oriented, combining existing multi-agent orchestration, visual tool invocation, multi-round SFT, and reinforcement learning techniques. The distinctive methodological novelty beyond these established components remains unclear. The authors should better clarify the core technical contributions of ShopEye.

  2. Dataset construction and annotation reliability. Although ShopEye-Bench provides a detailed annotation taxonomy, its construction relies heavily on an internal virtual try-on pipeline, automated annotation, and synthetic anomaly augmentation. The authors should provide stronger evidence of dataset diversity, independent annotation reliability, and train–test independence. The relationship between the reported 5K SFT, 12K RL, and 1K test samples also requires clarification.

  3. Fairness of the main comparisons. Table 1 shows an unusually large performance gap between ShopEye-Trainer and existing models. However, ShopEye-Trainer receives dedicated task-specific SFT and RL, while many baselines appear to be evaluated without equivalent adaptation. Comparisons under matched training data, optimization budgets, and tool-call limits are needed to establish whether the gains originate from the proposed methods rather than task-specific training.

  4. Dominance of EG in the ablation study. Table 2 shows that EG alone achieves 76.75 Overall F1, while combining LGDV and EG only marginally improves it to 76.83. Moreover, the proposed evidence-weighted reward improves over the standard RL reward baseline by only 1.88 points. These observations raise questions about the necessity of LGDV and the independent contribution of the proposed reward design.

  5. Numerical inconsistencies and experimental reproducibility. Section 6.2 reports an Overall F1 of 77.54 for ShopEye-8B, whereas Table 1 reports 76.83. Additionally, some reported F1 scores appear inconsistent with their corresponding precision and recall values. The authors should carefully verify the experimental results and clarify the evaluation protocol.

  6. Figure quality and presentation concerns. Figures 7 and 10 contain dense graphical elements and excessively small text, making the technical details difficult to inspect. Their highly stylized presentation also raises questions about figure preparation and verification. The authors should provide clearer, professionally prepared scientific diagrams, ensure that all depicted components accurately reflect the implementation, and clarify any generative-AI assistance if applicable.

Questions for authors

Please see the weaknesses’ part.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.