acceptodds
Withdrawn from ICLR 2027

ADClaw: Evidence-Grounded Active Inspection for Industrial Anomaly Detection and Reasoning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Industrial anomaly detection is essential for automated industrial inspection. Although multimodal large language models offer strong visual and semantic reasoning capabilities, existing methods often rely on a single observation or use tool outputs as transient feedback. This limits their ability to retain evidence across inspection steps and connect intermediate findings to final diagnoses. We introduce ADClaw, an evidence-grounded agent that formulates industrial anomaly diagnosis as active evidence construction and hypothesis verification. Its Structured Normality Prior integrates object–part semantics, explicit normality rules, and multi-scale normal visual memory to identify inspection targets and select suitable normal references. Guided by this prior, a persistent, programmable Inspection Workbench organizes heterogeneous tool outputs into a structured evidence state, allowing intermediate findings to be traced, composed, and reused throughout inspection. Hypothesis-Guided Evidence Reasoning assesses the sufficiency and consistency of this accumulated evidence to decide whether to acquire further observations, revise the current hypothesis, or finalize the diagnosis. These components work together in a hypothesize–inspect–reason–revise loop, turning discrete tool calls into a stateful inspection process in which each action addresses an explicit evidence requirement. Evaluations on 2D and 3D industrial anomaly detection benchmarks and MMAD demonstrate competitive detection and localization performance while enabling evidence-grounded defect understanding and traceable diagnostic explanations linking intermediate findings to final decisions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Industrial Anomaly Detection, Visual Anomaly Detection, Multimodal Large Language Models, Vision-Language Agents, Anomaly Localization

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

This paper proposes ADClaw, an agentic workflow tailored for 2D and 3D industrial anomaly detection. To enable evidence-grounded detection, the authors design an iterative, tool-callable pipeline integrated with an anomaly knowledge base. Empirical evaluations show that ADClaw achieves strong performance across several benchmarks compared to standard non-agentic baselines.

Critical strengths
  1. This paper provides a comprehensive and extensive empirical evaluation across a wide range of benchmarks, covering 2D, 3D, and multimodal industrial anomaly understanding scenarios.

  2. The proposed inspection workbench effectively minimizes redundant visual computations through stateful evidence resue, demonstrating benefits in reducing inference latency and computational overhead during multi-step inspection.

Critical weaknesses
  1. Unclear problem formulation and motivation
  • The paper introduces several components, including structured normality priors, persistent evidence management, and hypothesis-guided evidence reasoning. However, it remains unclear what specific limitation(s) is(are) solved from this architecture, and why these components are required to address this(those) limitation(s).

  • Overall, the framework appears to be an ad-hoc integration of an agentic workflow evaluated on public benchmarks. Due to the lack of precise problem formulation, it is difficult to identify what fundamental challenge is solved or the key scientific insights gained from this study.

  1. Unfair comparison with baselines
  • Table 1 and Table 2 compare ADClaw, which performs iterative, tool-augmented inspection, against baselines that operate in a single-pass setting without comparable tool-use capabilities. Consequently, the reported performance differences may reflect the effectiveness of whether tool usage is available or not.
  1. Insufficient Clarity
  • The number of normal samples used for inference is a crucial experimental factor in industrial anomaly detection. However, the evaluation protocol does not specify how many normal samples are utilized per inference. In other words, this missing detail makes the comparison with baselines ambiguous.

  • In Section 3.4, the authors describe Hypothesis-Guided Evidence Reasoning (HER). Although s̃ₑ is referred to as a calibrated score, the paper does not explain how this anomaly score is calibrated. Furthermore, key details regarding the predefined tool weight wew_{e} are missing, specifically, what values are used and how they vary across datasets. This is important because wew_{e} could critically affect performance. Lastly, the process of acquiring qeq_{e} is left completely unspecified. Since the authors highlight HER as one of their core contributions, these omissions raise significant concerns regarding clarity and reproducibility.

  • Table 5 compares the performance achieved under varying tool-calling setups. However, details regarding these setups are missing, making it difficult to determine which specific components actually drive the performance gains.

  1. Unclear Methodological Novelty
  • While the individual components are not strictly identical, both AnomalyClaw and ADClaw rely on heavily overlapping concepts that perform substantially similar functions—such as tool-based evidence usage, anomaly score calibration, candidate localization, iterative inspection loops, anomaly knowledge base (in case of AnomalyClaw, this uses RAG), and analytical capabilities. Because these core mechanisms serve equivalent roles, it remains difficult to discern what specific problem ADClaw aims to solve or what distinct advantages it offers over AnomalyClaw. Beyond minor technical variations, the underlying architectural philosophy appears virtually the same.
Questions for authors
  1. I noticed 3D dataset in the experiment while the main text mostly uses standard 2D illustrations. Is there anything special in how the agent handles 3D, or does the general reasoning pipeline work as-is ?
Reviewer_u5JhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ADClaw, a tool-augmented agent framework for industrial anomaly detection and understanding. Its Structured Normality Prior integrates object–part semantics, normality rules, and normal visual memory. A persistent, programmable Inspection Workbench stores tool outputs, their spatial scopes, and evidence dependencies. Hypothesis-Guided Evidence Reasoning then assesses the sufficiency and consistency of the accumulated evidence to determine whether to continue inspection, revise the current hypothesis, or produce a diagnosis. Experiments on four 2D anomaly detection datasets, two 3D anomaly detection datasets, and MMAD demonstrate the effectiveness of ADClaw.

Critical strengths
  1. The motivation and overall framework design are clear. The structured, reusable evidence representation aligns well with practical needs in industrial anomaly detection.
  2. The main experiments are comprehensive, and the quantitative results show substantial improvements of ADClaw over previous methods on industrial anomaly detection tasks.
  3. The ablation studies are complete and demonstrate the contribution of each mechanism.
Critical weaknesses
  1. The contribution is relatively limited. Compared with AnomalyClaw, the incremental advances of ADClaw primarily concern evidence management and verification procedures.
  2. Although A.2 states that the GPT-5.5 and AnomalyClaw results in the main MMAD experiments were obtained on a subset of the test data, this disclosure alone does not resolve the lack of comparability between the results of ADClaw and those of AnomalyClaw and other models.
  3. Limited localization performance. On MVTec 3D-AD, ADClaw performs substantially worse than several existing baselines on pixel-level metrics (P-AUROC and AUPRO). This suggests that the framework benefits anomaly detection more than anomaly segmentation.
Questions for authors
  1. This work is closely related to AnomalyClaw, as both use normal references and multiple rounds of tool calls to verify anomaly hypotheses. Why is AnomalyClaw not included in the main 2D and 3D anomaly detection experiments under the same experimental settings?
  2. Are the semantic knowledge and rule-based priors manually specified, or are they constructed by LLMs from normal reference samples? What is the cost of constructing these priors, and how is generalization to unseen anomaly types ensured?
  3. How does ADClaw maintain efficient inference? Its iterative planning and execution of multiple tools for evidence acquisition and verification appear computationally expensive. What is its end-to-end inference efficiency for anomaly detection?
  4. The paper sets inspection-step budgets according to the dataset type. If the budget is exhausted and no hypothesis satisfies the evidence acceptance criteria, what does ADClaw output?
Reviewer_APC4OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

ADClaw combines a frozen Qwen3.5-9B planner with object-part normality rules, normal visual memories, and a persistent programmable inspection workspace. Hypothesis-guided verification checks accumulated tool evidence before locking diagnostic outputs and generating explanations. Evaluation covers four 2D detection datasets, two XYZ-based 3D datasets, and the seven MMAD tasks.

Critical strengths
  1. The substantive system idea is to make inspection evidence persistent, addressable, and reusable across hypotheses, and to check its sufficiency before issuing a diagnosis. This addresses a real weakness of tool-using inspection agents: repeated tool outputs can remain disconnected observations rather than an accumulated basis for a decision. The combination of a normality prior, inspection workspace, and hypothesis verification is a coherent contribution even though the underlying models are frozen and existing tools supply the measurements.

  2. Table 5 supports value beyond ordinary tool access. With the backbone, tools, normal references, preprocessing, and evaluation protocol held fixed, average MMAD accuracy rises from 78.73 percent for a fixed pipeline to 80.32 for adaptive calling and 82.06 for ADClaw. This supports the complete inspection strategy, although its contribution is not isolated from inspection-budget differences.

Critical weaknesses
  1. The practical value of multi-step inspection is not established by the accuracy gains alone. Table 5 improves MMAD by 3.33 percentage points over the fixed pipeline and 1.74 over adaptive calling, but reports no corresponding end-to-end latency, throughput, or planner-token cost. The 2D inspection budgets allow up to 12 or 18 steps, without showing average realized steps or whether the compared strategies receive equal tool-call and planning budgets. The gains could therefore reflect additional inspection effort as well as better evidence organization. Figure 5 does not resolve this trade-off: its fixed-trace replay compares reuse against no reuse, reduces tool-stage time from 5.61 to 4.19 seconds, and explicitly excludes Qwen planning and subsequent postprocessing. It demonstrates avoided recomputation, not the cost of choosing an agent over a fixed pipeline. Multi-step diagnosis could be valuable for difficult-case review, but its suitability for production inspection depends on the intended workflow and acceptable latency. Without an accuracy-cost comparison, the paper does not establish when the agent's added complexity is worthwhile.

  2. The novelty resides in the inspection-specific organization and verification of evidence, since persistence and programmatic tool composition are established ideas. Section 2 already cites CodeAct's variable reuse and SpatialClaw's stateful kernel, while normal feature matching follows PatchCore. The paper therefore needs to make clear what its evidence representation and acceptance rules enable beyond a stateful visual agent with cached outputs. Table 5 is encouraging, but the adaptive-calling control does not establish that a simpler stateful cache with the same inspection budget would be insufficient. This comparison would locate the contribution at the level of the proposed evidence mechanism rather than generic persistence.

  3. The detection benchmarks do not establish the generality of the system's added value. Table 1 combines DINOv3 ViT-L/16 at 448 resolution with the inspection framework, whereas PatchCore uses Wide-ResNet50-2 at 256. Appendix C.3 separates hypothesis acceptance from dataset-specific scoring branches, and Appendix E states that Geometry3D determines the illustrated Real3D-AD object scores. The controlled MMAD gains support diagnostic QA, but the strong 2D and 3D scores alone cannot show that the same evidence-management principle improves detection and localization. A detector-only control with identical features and memories would test this broader value claim.

  4. The system provides an auditable record more clearly than a validated diagnostic explanation. Appendix A.2 evaluates description and analysis through multiple-choice accuracy, while Appendix E presents ten qualitative cases with edited summaries. Locking predictions before explanation generation is useful, but does not ensure that each explanatory claim follows from the recorded evidence. Since verifiable explanations are a central reason to introduce the workspace and reasoning machinery, unsupported-claim rates on unedited reports would be a direct measure of the proposed system's value.

Questions for authors

I consider this a borderline system contribution. The QA gains and reduced recomputation support empirical usefulness, but practical value remains uncertain because the full cost of multi-step inspection is unreported. The decision should turn on whether explicit evidence management adds enough beyond simpler stateful execution to justify its cost in a defined inspection workflow, and whether its explanations are reliably supported.

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

Related papers

Discussion (0)

Sign in to comment.