acceptodds
Withdrawn from ICLR 2027

AD-VRPD: ANOMALY DETECTION VIA VERIFIABLE ROLLOUT-GUIDED SELF-DISTILLATION

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Zero-shot visual anomaly detection requires identifying and localizing defects without reference images at inference, yet generic anomaly queries leave the expected normal state and relevant defects unspecified. We address this ambiguity through diagnostic-text-conditioned inspection and adapt MMR-AD by pairing images with reusable descriptions of normality and possible defects. Explicit criteria clarify the task, but accurate localization remains challenging, particularly for small defects. We propose AD-VRPD, a verifiable rollout-guided on-policy self-distillation method that transfers spatial guidance from a frozen teacher viewing GT-marked images to a student viewing unmarked images. A task verifier regulates this supervision by reconstructing distillation targets when the teacher's guidance conflicts with verified rollout outcomes. Across four datasets held out from detector training, AD-VRPD improves localization over its pretrained backbone and vanilla self-distillation, enabling a compact multimodal detector to achieve competitive performance against larger and proprietary models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
anomaly detection, knowledge distillation, multimodal learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_FNjXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes AD-VRPD, a verifiable-reward-guided self-distillation method for industrial anomaly detection. It first uses Qwen3.7-plus to construct diagnostic descriptions of normal states and potential defects for different defect types. A frozen teacher receives ground-truth-box-marked images, while the student operates on unmarked images, enabling privileged-view online self-distillation. The method employs trajectory-level rewards computed from ground-truth anomaly labels and bounding boxes to determine whether to use standard distillation or a reconstructed target based on the consistency between the teacher's sampled-token guidance and the reward signal. The detector is trained on ten datasets and evaluated on four additional datasets for image-level anomaly decisions and bounding-box localization. The reported results show improvements over the evaluated baselines.

Critical strengths
  1. The paper explicitly describes normal states and potential defects through diagnostic text, providing clearer inspection criteria for industrial anomaly detection. This design helps address the lack of specific criteria in generic anomaly prompts and offers a practical formulation for condition-conditioned anomaly detection.

  2. The paper uses trajectory-level rewards to assess whether the teacher's guidance aligns with the quality of the generated response, reconstructing the distillation target when conflicts arise. This design addresses the potential unreliability of privileged teacher supervision and offers a potentially useful mechanism for online self-distillation.

Critical weaknesses
  1. The diagnostic texts are constructed using examples and annotations from the test sources, and anomalous test images are paired with descriptions selected according to their ground-truth defect types. Consequently, the evaluation is not fully independent of target-domain information and does not establish the method's ability to detect anomalies autonomously under unknown defect types. This limitation directly undermines the paper's claims regarding cross-dataset generalization.

  2. The method uses a single trajectory-level reward to regulate distillation across the entire output sequence, although anomaly classification and bounding-box localization can have different levels of correctness. For example, a model may correctly identify an anomaly but produce poorly localized boxes, while the current mechanism does not explicitly distinguish these error types. Consequently, the reward may provide insufficiently precise supervision for different output components, and the paper does not adequately establish whether trajectory-level target reconstruction is preferable to more targeted supervision strategies.

Questions for authors
  1. The diagnostic texts are constructed using all 14 datasets, including the test sources, and anomalous test images are assigned descriptions according to their ground-truth defect types. Could the authors clarify how this setup affects the independence of the evaluation and provide results using diagnostic texts constructed solely from training data, without selecting descriptions based on test labels?

  2. At test time, anomalous images are paired with diagnostic texts selected according to their ground-truth defect types, whereas the defect type is generally unknown in practical industrial inspection. How should the model select an appropriate diagnostic text without access to ground-truth information, and to what extent do the reported results depend on this defect-type-conditioned input setting?

  3. The method uses a single trajectory-level reward to regulate distillation across the entire output sequence, although anomaly classification and bounding-box localization may have different levels of correctness. Could the authors provide ablations demonstrating that this unified reward signal improves distillation beyond simpler strategies that supervise classification and localization separately?

Reviewer_t7DpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies reference-free visual anomaly detection under a diagnostic-text interface: an unmarked query image is paired with a reusable description of the expected normal state and possible defect types, and the model outputs an anomaly decision plus bounding boxes. The authors adapt MMR-AD, construct shared defect-type descriptions with an LLM, and evaluate on a dataset-disjoint split (10 training sources, 4 test sources: MVTec AD, VisA, MVTec 3D-AD, MPDD). The method, AD-VRPD, adapts privileged-view on-policy self-distillation: a frozen teacher initialized from the same weights sees GT-box-marked images while the student sees unmarked ones, and a verifier-gated target reconstruction replaces teacher supervision at token positions whose direction conflicts with a trajectory-level reward. The headline comparison is a three-row table: [email protected] goes 34.2 (base Qwen3-VL-8B) -> 45.4 (vanilla OPSD) -> 46.1 (AD-VRPD); the paper attributes most of the gain to OPSD and +2.0/+0.7 points at IoU 0.1/0.5 to the proposed reconstruction.

Critical strengths

1. A clearly documented diagnostic-text protocol with unusually candid disclosure of its limitations. The formulation (normal-state plus defect-type text, dataset-disjoint detector split) is a reasonable benchmark contribution, and the paper states in three places (Sec. 3.2, App. C, Limitations) that test-source examples and annotations enter text construction and that anomalous queries receive GT defect-type descriptions. This transparency lets the community judge the protocol rather than be misled by it.

2. An honest attribution of where the gains come from. The paper explicitly reports that privileged-view OPSD accounts for most of the localization improvement and that verifier-guided reconstruction adds only a small margin (Sec. 5.3, Conclusion), and it reports a reversed result on the logical-anomaly subgroup (App. A). Reporting nuanced or negative findings rather than only the favorable aggregate is a genuine service to the field.

Critical weaknesses

1. The proposed method's contribution is insufficiently isolated. Table 2 compares only the base model, OPSD, and AD-VRPD. Base-to-OPSD conflates training on 49k samples, privileged teacher input, and on-policy sampling. The incremental AD-VRPD gain is only +0.7 [email protected], without separating gating from target reconstruction. Key ablations (mask-only, inverse gate, reconstruction-only, and original teacher) and matched SFT/GRPO baselines are missing. These controls are particularly important given OPDVR's [1] GRPD/inverse-gate experiments and IoU-PD's [2] strong SFT results. Related approaches such as RG-OPD [3], AR-OPD [4], ADOPD [5], and MOPDA [6] further motivate clearer methodological differentiation. Without these controls, the benefit of the proposed mechanism remains unclear.

2. The incremental gain lacks robustness analysis and convincing localization evidence. The +0.7 [email protected] improvement is reported from a single seed, without variance or confidence intervals. Both the verifier threshold (τ=0.1) and checkpoint selection (validation [email protected]) emphasize permissive overlap, where the gain is larger (+2.0). No threshold sensitivity or verifier-conflict analysis is provided. Moreover, Base-to-OPSD box recall increases by 15.5 points versus only 2.2 in precision, leaving the effects of prediction coverage versus localization accuracy insufficiently disentangled. Per-dataset results are also missing, despite the logical-anomaly subgroup showing a performance reversal relative to the base model (Appendix A).

3. Table 2 appears inconsistent with the stated evaluation protocol. Under Appendix B's micro-averaged definitions, the total number of predicted boxes should be invariant to the IoU threshold. However, OPSD's reported precision and recall imply approximately 4,025 predicted boxes at IoU 0.1 versus 3,839 at IoU 0.5, a discrepancy too large to explain by rounding. The authors should verify these results or clarify the evaluation procedure.

References

[1] W. Lin et al. On-policy Distillation with Verifiable Reward. arXiv:2608.24696, 2026. [2] X. Zhu et al. IoU-PD: IoU-Aware Privileged Distillation for Visual Grounding with Multimodal Large Language Models. arXiv:2607.15732, 2026. [3] M. S. Akhondzadeh et al. Reward-Gated On-Policy Distillation. arXiv:2607.04037, 2026. [4] W. Zhang. Beyond Absolute Imitation: Anchored Residual Guidance for Privileged On-Policy Distillation. arXiv:2606.10385, 2026. [5] J. He et al. ADOPD: Reference-Privileged On-Policy Distillation for MLLM-Based Industrial Anomaly Detection. arXiv:2608.09789, 2026. [6] S. Chen et al. MOPDA: Mixed-Trajectory On-Policy Distillation for Language-Guided Industrial Anomaly Detection. arXiv:2607.18850, 2026.

Questions for authors
  1. Missing baselines. Table 2 conflates additional training, privileged teacher input, and on-policy sampling. Under matched data, backbone, and comparable budgets, could you report (i) SFT, (ii) GRPO/GRPD using your verifier, and (iii) OPSD with an unmarked-image teacher? Please also clarify the methodological differences from IoU-PD [2], RG-OPD [3], AR-OPD [4], ADOPD [5], and MOPDA/OPD-IAD [6].
  2. Component isolation. The paper motivates target reconstruction over masking without directly comparing them. Could you evaluate gate+mask, gate+reconstruction, reconstruction without gating, and inverse gating? Please also report the conflict rate (fraction of tokens with ft<0f_t < 0) and sensitivity to τ\tau.
  3. Significance and selection. The +0.7+0.7 [email protected] gain comes from a single seed, with checkpoints selected using [email protected]. Could you report mean/std across at least three seeds, paired confidence intervals, and results when selecting checkpoints by validation [email protected]? Does the incremental gain remain reliable?
  4. Localization and result consistency. Could you report predicted boxes per image, matched-box IoU distributions, and per-dataset results to distinguish improved localization from increased detection coverage? Appendix A shows a reversal relative to the base model on logical anomalies. Additionally, Table 2's OPSD precision/recall values imply approximately 4,025 predicted boxes at IoU 0.1 versus 3,839 at IoU 0.5, seemingly inconsistent with Appendix B's micro-averaged protocol. Could you clarify?

A matched ablation study covering SFT, original/privileged teachers, gate-only, and full AD-VRPD, alongside a GRPO baseline and multiple seeds, would substantially strengthen the evidence. These results would materially affect my assessment.

Reviewer_adKgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper identifies two issues in VLM-based anomaly detection: • underspecified prompts without context on what an anomaly may be • Most anomalies are very small (<1% of the image) defects which VLMs find challenging to localize (lines 51-52) The authors provide text describing the normal state and anomalous defects to their VLM prompts and then propose AD VRPD as a verifier guided on policy self distillation method for anomaly detection and localization to address these challenges. AD-VRPD has a teacher network that trains using images with anomaly-localizing bounding boxes and a student that receives image without bounding boxes.

Critical strengths

The paper references relevant motivations in anomaly detection: Localizing small anomalies or defects is challenging for MLLMs/VLMs. Users may also under-specify anomaly detection prompts to models in real-world settings.

Critical weaknesses

Motivations for proposed methods are unexplained or anecdotal: The paper uses examples of VLM outputs referring to red boxes to motivate the proposed teacher supervision. However, these examples are anecdotal motivation and the paper does not measure how often these errors occur or whether AD-VRPD reduces them.

Test set information involved in training and evaluation: The VLM prompts are constructed using examples from the test set (line 45-46), this is acknowledged as a limitation but the experiments do not then isolate whether gains are due to involving test set information or whether reported performance holds when prompts are constructed without test-set information.

AD-VRPD performs worse than the Qwen baseline on logical anomalies. Its added gains over OPSD are modest, and it ranks third on both box-level F1 metrics.

The paper is generally hard to follow: for example it does not specify that ground truth bounding boxes are red, this is only mentioned in the caption of figures or figures themselves (Figure 2 and Figure 5) and it is left to the reader to infer what "red box" in the text refers to (lines 70-71) . Several abbreviations used such as GT, ADOPD which are unclear and unreferenced. It is left to the reader to assume they refer to ground truth or are method names. Several models and datasets used in the paper are also not cited.

Questions for authors

Compare against baselines fine-tuned on the same training split, e.g., SFT or GRPO Provide results using text descriptions not derived from the test set or experiments isolating whether this contributes to gains in performance Can the novelty compared to OPDVR be clarified? Can an explanation be provided for why AD-VRPD is worse than the Qwen baseline for logical anomalies?

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

Related papers

Discussion (0)

Sign in to comment.