acceptodds
Withdrawn from ICLR 2027

MOPDA: Mixed-Trajectory On-Policy Distillation for Language-Guided Industrial Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Large vision-language models (LVLMs) have recently shown strong potential for industrial anomaly detection (IAD) by providing image-level anomaly judgments and interpretable defect reasoning. However, reliably translating generated language judgments into precise pixel-level anomaly localization remains challenging. To address this, we propose Mixed-Trajectory On-Policy Distillation for Language-Guided Industrial Anomaly Detection (\MODPA), the first framework to introduce on-policy self-distillation into LVLM-based IAD. Specifically, \method introduces Mixed-Trajectory Supervision, which combines student-generated on-policy trajectories with teacher-generated evidence-conditioned trajectories under a shared token-level distillation objective for fine-grained judgment learning. To connect the learned judgment with dense anomaly perception, we further introduce Language-guided Visual Anchoring. The final judgment is used as a compact semantic condition to construct image-specific normal and abnormal anchors, which are contrasted with dense visual features to produce the anomaly map. In this way, language guides localization without directly determining the pixel-level response. Experiments on five IAD benchmarks show that \method outperforms the evaluated LVLM-based baselines on most detection, localization, and judgment metrics, while remaining competitive with CLIP-based methods.

Primary area:
reinforcement learning
Keywords:
RL, OPSD, Anomaly Detection

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_7483OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work focuses on zero-shot industrial anomaly detection. It develops Mixed-Trajectory Supervision that combines student-generated on-policy trajectories with evidence-conditioned teacher-generated trajectories. It also introduces Language-guided Visual Anchoring to guide anomaly localization using the final language judgment. The paper reports its main experimental results on five IAD benchmarks.

Critical strengths

The paper explores a language-guided Contrastive Heatmap Head for segmenting anomalous regions, offering a reasonable and task-driven design.

The main results are evaluated using a comprehensive set of metrics, including image- and pixel-level AUROC, AP, and F1, as well as QA/Acc. for language judgments.

Critical weaknesses
  1. Limited Methodological Novelty

The two main contributions of this paper appear to be largely incremental, and their substantive novelty over existing methods has not been sufficiently established.

  • Mixed-Trajectory Supervision: The proposed approach appears to primarily extend existing On-Policy Distillation (OPD) by incorporating teacher-generated trajectories while retaining the conventional token-level distillation objective. The paper does not provide a theoretical justification for why mixed-trajectory supervision should outperform standard OPD, nor does it offer sufficiently in-depth empirical analyses to explain the underlying mechanisms responsible for its effectiveness.
  • Language-guided Visual Anchoring: Using special semantic tokens generated by large vision-language models (LVLMs) to guide pixel-level segmentation has been extensively explored in prior work, such as LISA (Lai et al., 2024) and PixelLM (Ren et al., 2024). Although the proposed method introduces Judgment Reforward and normal/abnormal anchors, the authors do not sufficiently discuss the substantive differences between their approach and existing language-guided segmentation paradigms. Furthermore, no controlled comparisons with representative methods in this area are provided. Consequently, the independent novelty of this component and its advantages over existing approaches remain insufficiently demonstrated.
  1. Insufficient Experimental Validation

The experimental evaluation is not sufficiently comprehensive, particularly regarding comparisons with relevant prior work and analyses of the proposed mechanisms.

  • Missing Important Baselines: The paper omits comparisons with several important and relevant methods, including IAD-R1 (Li et al., 2026), a representative LVLM-based industrial anomaly detection and reasoning method, as well as competitive zero-shot anomaly detection approaches such as Bayesian Prompt Flow Learning for Zero-Shot Anomaly Detection (Qu et al., 2025, CVPR) and Anomaly-OV (Xu et al., 2025, CVPR). The absence of these comparisons weakens the evidence supporting the claimed performance advantages.
  • Lack of In-depth Analysis: Most of the current experiments focus on final anomaly detection performance, with limited investigation into the proposed learning mechanisms. In particular, the paper lacks direct quantitative evaluations of the quality of the generated chain-of-thought (CoT) reasoning. Moreover, there are insufficient analytical experiments to explain why mixed-trajectory sampling enables the student model to learn more effectively. Such analyses are important for validating the central motivation and effectiveness of the proposed method.

References

  1. Lai, X., et al. (2024). LISA: Reasoning Segmentation via Large Language Model. CVPR 2024.
  2. Ren, Z., et al. (2024). PixelLM: Pixel Reasoning with Large Multimodal Model. CVPR 2024.
  3. Li, Y., et al. (2026). IAD-R1: Reinforcing Consistent Reasoning in Industrial Anomaly Detection. AAAI 2026.
  4. Qu, Z., et al. (2025). Bayesian Prompt Flow Learning for Zero-Shot Anomaly Detection. CVPR 2025.
  5. Xu, J., et al. (2025). Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models. CVPR 2025.
Questions for authors

What substantive methodological advances do mixed-trajectory supervision and language-guided segmentation offer over existing approaches, and what deeper analyses could explain why these components are effective?

Reviewer_xcvkOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MOPDA (Mixed-Trajectory On-Policy Distillation for Language-Guided Industrial Anomaly Detection), a framework that combines LVLM reasoning with dense anomaly localization. The paper identifies the difficulty of translating language-based anomaly judgments into pixel-level anomaly maps and proposes two main components: (1) Mixed-Trajectory Supervision, which combines student-generated on-policy trajectories with teacher-generated evidence-conditioned trajectories under token-level distillation; and (2) Language-guided Visual Anchoring, which uses the final anomaly judgment as semantic guidance to construct normal/abnormal anchors for dense localization while keeping pixel-level prediction grounded in visual features. The framework is evaluated on five industrial anomaly detection benchmarks, comparing against CLIP-based and LVLM-based baselines. The reported experiments show improvements on detection, localization, and language judgment metrics over evaluated LVLM baselines.

Critical strengths

The proposed Language-guided Visual Anchoring is conceptually appealing. Instead of directly using language tokens or reasoning attention as localization signals, the method uses the final judgment to construct semantic anchors and performs dense scoring with visual features. This design avoids the potential instability caused by directly relying on language-side outputs.

Critical weaknesses
  1. The central contribution claims to introduce the first on-policy self-distillation framework for LVLM-based IAD and proposes mixed trajectories combining student rollouts and teacher rollouts.
    However, the conceptual difference from existing methods such as:
  • privileged-context distillation;
  • reference-privileged on-policy distillation;
  • self-distilled reasoning; is not fully convincing. The paper already acknowledges related privileged-context self-distillation and EMA teachers.
    The main question is: What fundamentally changes because the teacher trajectory is mixed with student trajectory? The current ablation shows MOPDA improves over OPSD, but the gain is relatively small on some metrics. More analysis is needed to demonstrate that the improvement comes from the mixed trajectory mechanism rather than simply stronger teacher supervision. Suggested additional analysis:
  • Compare teacher-only trajectory distillation vs mixed trajectory;
  • Vary λ systematically;
  • Analyze trajectory diversity and error correction.
  1. The localization contribution may rely heavily on additional supervision rather than the proposed semantic mechanism The localization branch uses:
  • image-level anomaly supervision;
  • pixel-level mask supervision;
  • focal loss;
  • Dice loss;
  • anchor separation loss.

Therefore, the improvement in localization may partially come from adding a supervised segmentation head rather than from the language-guided anchoring idea itself. The key scientific question is: Does language judgment actually provide useful localization information, or is it merely an auxiliary condition for a supervised anomaly segmentation module?

The oracle judgment experiment is helpful, but stronger controlled experiments are needed.

Questions for authors
  1. How does MOPDA differ fundamentally from existing privileged-context on-policy distillation methods?

  2. How sensitive is MOPDA to the trajectory mixing ratio λ? The paper reports λ=0.8, but the contribution of this design would be clearer with a sensitivity analysis.

  3. Can the authors separate the contribution of language guidance from the supervised localization branch? For example:

  • remove language conditioning while keeping identical localization losses;
  • use random judgments instead of generated judgments;
  • compare against purely visual anchor learning.
  1. The teacher requires privileged defect evidence during training. How expensive is this annotation requirement in realistic industrial scenarios? The paper should discuss whether the method remains practical when detailed defect reasoning references are unavailable.
Reviewer_F6WGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MOPDA for language-guided industrial anomaly detection. It combines Mixed-Trajectory Supervision with Language-guided Visual Anchoring. MTS distills student-generated and privileged teacher-generated trajectories. LVA uses the final binary judgment to construct normal and abnormal anchors for dense localization. Experiments on five benchmarks show promising results.

Critical strengths
  1. The paper addresses the challenge of connecting LVLM-based anomaly judgments with accurate pixel-level localization.

  2. The evaluation covers five industrial anomaly detection benchmarks. MOPDA achieves strong results on several datasets.

  3. Useful experimental analyses. The paper compares SFT, GRPO, OPSD, and MOPDA.

Critical weaknesses
  1. Limited novelty of Mixed-Trajectory Supervision [1–3]. GKD [1] already combines student-generated and non-student trajectories through stochastic sampling and token-level divergence. OPSD [2] introduces privileged-context self-distillation, while ADOPD [3] applies related techniques to industrial anomaly detection.

  2. Insufficient novelty of visual anchoring [4, 5]. AG-VAS [4] already employs normal/abnormal semantic anchors and semantic-pixel alignment. VisualAD [5] uses learnable normal/abnormal tokens for dense localization. MOPDA mainly adds judgment-conditioned reforwarding, but the judgment contains only binary class information.

  3. Incomplete and potentially unfair comparisons [4–6]. Table 1 omits AG-VAS [4], despite its close methodological relevance. MOPDA uses full-parameter Qwen3-VL-8B fine-tuning and pixel-level supervision, while several baselines use different architectures or weaker supervision.

References:

[1] Agarwal et al. On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes. ICLR, 2024.

[2] Zhao et al. Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models. ICML, 2026.

[3] He et al. ADOPD: Reference-Privileged On-Policy Distillation for MLLM-Based Industrial Anomaly Detection. arXiv:2608.09789, 2026.

[4] Qu et al. AG-VAS: Anchor-Guided Zero-Shot Visual Anomaly Segmentation with Large Multimodal Models. CVPR, 2026.

[5] Hou et al. VisualAD: Language-Free Zero-Shot Anomaly Detection via Vision Transformer. CVPR, 2026.

[6] Jin et al. Reasoning-Driven Anomaly Detection and Localization with Image-Level Supervision. CVPR, 2026.

Questions for authors
  1. What is the precise algorithmic difference between MTS and GKD or OPSD? Can the authors provide a same-backbone GKD-style baseline with identical supervision and localization components?

  2. Why is binary judgment-conditioned reforwarding necessary beyond existing anchor methods? Can the authors compare predicted judgments, random binary tokens, fixed tokens, and judgment-free reforwarding under matched computation?

  3. Why is AG-VAS missing from Table 1? Can the authors provide comparable supervision settings and explain the weaker WFDD and KSDD2 results?

  4. Does MOPDA improve localization robustness under language perturbations? Can the authors compare heatmap consistency with AG-VAS and ReAL? How often does localization remain correct despite an incorrect language judgment?

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

Related papers

Discussion (0)

Sign in to comment.