acceptodds
Withdrawn from ICLR 2027

IndusAgent: Agentic Reinforcement Learning for Cross-Category Industrial Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Small defects in unfamiliar product categories can be difficult to distinguish from benign appearance variations in a global view. Additional reasoning over the same view may leave this ambiguity unresolved. We introduce IndusAgent, a Qwen3-VL-8B inspection policy trained through supervised fine-tuning and tool-augmented reinforcement learning for cross-category industrial anomaly detection. The policy selects among six tools for cropping, enhancement, measurement, within-image comparison, texture analysis, and external visual opinions from Qwen3-VL-Max, then integrates returned observations into a structured diagnosis. We construct Indus-CoT with 3,000 trajectories from 26 Real-IAD categories, generated with actual tool feedback and screened against annotations after teacher generation. GRPO optimizes an accuracy-gated objective that conditions localization, type, and tool rewards, including call costs, on final diagnostic correctness while retaining an independent format reward. Under the common protocol, IndusAgent exceeds direct Max inference in mean balanced accuracy across five benchmarks (83.5% versus 79.4%). Across three training seeds, the gated policy achieves higher five-benchmark mean balanced accuracy than the weight-matched ungated control (83.5% versus 80.5%). Evaluation uses disjoint product categories and a fixed defect vocabulary. Inference requires no paired normal exemplar but can access the external prior API.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision Language Model, Agentic Reinforcement Learning, Industrial Anomaly Detection, Chain-of-Thought

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

The paper addresses cross-category industrial anomaly detection, where small defects on unfamiliar product types are hard to distinguish from benign appearance variations under a global view. The authors propose IndusAgent, a tool-augmented reinforcement-learning framework built on a Qwen3-VL-8B policy that learns to inspect before deciding. There are six tools: crop, enhancement, geometric measurement, within-image comparison, texture analysis, and an external visual prior queried from Qwen3-VL-Max.

Critical strengths

The framework is supported by a structured dataset that integrates global visual observations, high-resolution local patches, and expert normalcy priors to supervise industrial inspection trajectories.

IndusAgent actively utilizes and coordinates external tools during inspection, including dynamic region cropping, high-frequency feature enhancement and prior retrieval.

The agent's decision-making and tool usage are optimized via a gated reinforcement learning objective that jointly improves anomaly classification, localization accuracy, and reasoning capabilities.

Critical weaknesses

The core performance gains seem to heavily rely on standard visual prompt techniques and basic external tools (e.g., dynamic cropping, high-frequency filtering). The conceptual leap of wrapping these tools in an LLM agent framework for IAD is incremental, and it is unclear whether the complexity of the agentic workflow is strictly necessary compared to simpler, optimized visual prompting baselines.

There lacks a mathematically rigorous formulation of the reward function of Gated Reinforcement Learning. It is unclear how the reward balances pixel-level anomaly localization with categorical reasoning.

The authors evaluate IndusAgent against its base model (QN3VL8B) and general 72B parameter models. They fail to compare the framework against state-of-the-art specialist open-vocabulary and zero-shot IAD models (e.g., WinClip [1], AnomalyCLIP [2], or recent agentic frameworks like AgentIAD [3], AnomalyAgent [4]).

The paper does not provide a detailed comparison with existing agentic anomaly detection methods. How does the proposed approach differ from these methods [3,4], and what advantages does it offer over them?

[1] Jeong, Jongheon, et al. "Winclip: Zero-/few-shot anomaly classification and segmentation." IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023.

[2] Zhou, Qihang, et al. "Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection." International Conference on Learning Representations. Vol. 2024. 2024.

[3] Miao, Junwen, et al. "AgentIAD: Agentic Industrial Anomaly Detection via Adaptive Memory Augmentation." arXiv preprint arXiv:2512.13671 (2025).

[4] Zhang, Yi, et al. "AnomalyAgent: Training-Free Agentic Models for Zero-/Few-Shot Anomaly Detection." arXiv preprint arXiv:2605.30140 (2026).

Since the authors employ an LLM for anomaly detection, they should provide experiments to evaluate the generalization ability of the proposed method. However, the paper does not provide any such evaluation.

Questions for authors

Industrial anomaly detection is a well-studied problem. What specific limitations of industrial anomaly detection can be addressed by introducing LLMs or agents? Beyond presenting the work in a relatively well-structured manner, what meaningful contribution does the proposed method make to the field of industrial anomaly detection?

3 comments

Thank you for the detailed feedback on the agent's contribution, comparisons, and generalization.

Y1: What is the contribution of adaptive tool use compared with a trained fixed inspection pipeline?

A1: We appreciate the question about the learned policy's contribution. Our contribution is a learned inspection policy that chooses actions for each image and integrates feedback from their execution. Full reaches 83.4 balanced accuracy (BA), compared with 82.6 for the trained fixed Crop-plus-Prior policy, while using 0.17 versus 1.00 Prior calls per query (Table 1). Latency falls from 2.76 to 1.47 seconds, approximately 46.7%, and API charges fall by 84.2%. Table 2 reports all five benchmarks; the fixed policy is slightly better on MPDD, at 72.8 versus 72.6.

All trained controllers share Qwen3-VL-8B, accepted source images, supervision and update budgets, preprocessing, and evaluation. Prior-enabled configurations use the same Qwen3-VL-Max interface. The fixed controller retains SFT, the full R11 reward, and GRPO. It learns image-dependent crop coordinates and Prior focus, and executes one Crop and one Prior. Each controller receives feedback from its own executed actions.

Full can choose among six tools, whereas the fixed controller uses Crop and Prior. The comparison therefore includes both scheduling and the available tool set. The IAD-R1-style and AgentIAD-style adaptations can also differ in objectives or tool configurations. Base plus fixed Prior has no industrial training. This trained fixed-policy control evaluates an alternative with fixed tool use; tool-removal experiments measure sensitivity to tool availability.

Table 1. Controller performance and resources

BA is a percentage. Means equally weight MVTec-AD, MPDD, VisA, DTD, and SDD. Calls and resource figures are per query. Prior is the external Qwen3-VL-Max component; API charges exclude local computation. The supplementary Full result, 83.4, is reported separately from the manuscript result of 83.5.

PolicyMean BALocal / Prior callsSeconds/queryAPI CNY/query
IAD-R1-style without Prior79.20 / 01.070
IAD-R1-style with Prior82.00 / 12.580.0039
AgentIAD-style82.20.65 / 0.301.680.00115
Fixed Crop + Prior82.61 / 12.760.0038
Base + fixed Prior78.80 / 12.470.0037
Full83.40.78 / 0.171.470.0006

Table 2. Per-benchmark balanced accuracy (%)

PolicyMVTecMPDDVisADTDSDDMean
IAD-R1-style without Prior80.570.272.090.782.679.2
IAD-R1-style with Prior82.772.075.894.285.382.0
AgentIAD-style82.972.375.994.585.482.2
Fixed Crop + Prior83.272.876.594.785.882.6
Base + fixed Prior79.567.872.590.583.778.8
Full83.572.676.795.588.783.4

Y2: How is the reward formulated, how are its gates evaluated, and how is localization rewarded?

A2: Thank you for asking how the reward treats localization. The full reward is

R=C[1+0.8L+0.6T+0.15I−0.05N]+F.R=C[1+0.8L+0.6T+0.15I-0.05N]+F.

Here CC is final diagnostic correctness, LL localization reward, TT canonical-type agreement, II the successful-call and positive-margin-change indicator, NN admitted calls, and FF format reward. Correctness gates auxiliary credit and call cost; format reward remains independent. Localization uses continuous box IoU. The box-IoU threshold of 0.5 used to screen teacher trajectories serves a separate data-selection role. The tool-progress term uses confidence change following a successful observation as a proxy for progress.

We separate positive-credit gating from cost gating using

P=0.8L+0.6T+0.15I,P=0.8L+0.6T+0.15I,

K=0.05N,K=0.05N,

R00=C+P−K+F,R_{00}=C+P-K+F,

R10=C+CP−K+F,R_{10}=C+CP-K+F,

R11=C+CP−CK+F.R_{11}=C+CP-CK+F.

Y3: How does the method compare with specialist anomaly detectors such as WinCLIP and AnomalyCLIP?

A3: Thanks for suggesting comparisons with specialist detectors. Full improves both detection metrics over the two CLIP-based specialists on the same target sets: 90.1 AUROC / 83.4 BA, compared with 82.6 / 77.9 for WinCLIP and 86.1 / 81.2 for AnomalyCLIP (Table 4). The specialists are substantially faster, taking 0.08 and 0.09 seconds per query versus 1.47 for Full. This comparison shows an accuracy gain at higher inference latency.

Table 4. Specialists and native-style agents

All systems are evaluated on the same target sets. BA thresholds are individually selected on source data, and auxiliary supervision excludes the target product categories. AUROC uses continuous anomaly scores. Each system retains its native backbone and budget; external calls use that system's own service units, not Qwen3-VL-Max-equivalent units. Scores are percentages and means equally weight MVTec-AD, MPDD, VisA, DTD, and SDD. Resources are per query; API charges exclude local computation.

SystemAUROCBASeconds/queryExternal callsAPI CNY/query
WinCLIP82.677.90.0800
AnomalyCLIP86.181.20.0900
AnomalyAgent native-style90.884.14.802.400.0240
Full90.183.41.470.17 Prior0.0006

Y4: How does the method compare with existing agents, and what external visual information does it use?

A4: Thank you for bringing up existing agent baselines. AnomalyAgent native-style reaches 90.8 AUROC and 84.1 BA, 0.7 points above Full on both metrics (Table 4). Full takes 1.47 versus 4.80 seconds per query and incurs 0.0006 versus 0.0240 CNY in API charges. It retains closely comparable accuracy with substantially lower latency and API expenditure.

The common-backbone comparison in Table 5 also includes an adapted AgentIAD-style controller: Full reaches 83.4 BA versus 82.2, with 0.17 versus 0.30 Prior calls. Full combines an 8B policy with Qwen3-VL-Max as the query-conditioned Prior. Reference-free inference means that no paired normal exemplars are supplied; the Prior provides external visual information.

Table 5. Controller performance and resources

Trained controllers share Qwen3-VL-8B, accepted source images, supervision/update budgets, preprocessing, and evaluation. Prior-enabled rows use the same interface. Fixed retains SFT/R11/GRPO, learns crop coordinates and Prior focus, and executes one Crop and one Prior. Each controller receives feedback from its own actions. Full has six available tools; Fixed has two. Adapted methods can differ in objectives or tools. Base plus fixed Prior has no industrial training. BA is a percentage averaged equally over the five benchmarks named above. Calls and resources are per query, Prior is Qwen3-VL-Max, and API charges exclude local computation.

PolicyMean BALocal / Prior callsSeconds/queryAPI CNY/query
IAD-R1-style without Prior79.20 / 01.070
IAD-R1-style with Prior82.00 / 12.580.0039
AgentIAD-style82.20.65 / 0.301.680.00115
Fixed Crop + Prior82.61 / 12.760.0038
Base + fixed Prior78.80 / 12.470.0037
Full83.40.78 / 0.171.470.0006

Y5: What kind of generalization does the evaluation establish?

A5: We appreciate the request to clarify the generalization setting. The experiment tests transfer to product categories held out from industrial post-training. Target categories are excluded before trajectory construction, and the canonical defect vocabulary remains fixed. The result therefore concerns product-category transfer within that vocabulary. Novel defect taxonomies, foundation-model pretraining exposure, and changes in acquisition conditions are separate generalization questions beyond this protocol.

Reviewer_Ku26OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes IndusAgent, an agent framework for industrial defect detection that selectively uses analytical tools and visual assessments from an external vision-language model (VLM). The agent operates in two rounds: preliminary diagnosis and tool selection, followed by a final diagnosis that integrates the resulting observations. It first undergoes supervised fine-tuning (SFT) on 3,000 Indus-CoT trajectories containing feedback from actual tool execution. It is then trained with accuracy-gated GRPO, which applies auxiliary rewards for localization, defect type, and tool use only when the final diagnosis is correct. The full system achieves an average balanced accuracy of 83.5% across five benchmarks, exceeding direct inference with Qwen3-VL-Max by 4.1%p.

Critical strengths
  • The problem formulation, motivation, and proposed approach are clearly connected. Selecting additional observations and incorporating them into the final diagnosis is well suited to detecting localized defects that are difficult to assess from global image judgments alone.
  • The reward design is well motivated. Comparisons with an ungated reward formulation using identical weights, evaluated across three random seeds, support the effectiveness of the full gating scheme.
  • The paper reports efficiency alongside detection performance. Under a common evaluation protocol, the method achieves higher average performance than direct inference with Qwen3-VL-Max, with fewer calls and lower latency in the reported measurement setting. The evaluation also includes localization and defect-type metrics, as well as false positives on normal images.
Critical weaknesses
  • Limited information about the teacher model hinders reproducibility of the generated dataset. The teacher's data-generation settings are not clearly specified. Although the filtering of final outputs is described, the information provided is insufficient to assess the correctness of the teacher's reasoning or potential biases in the generated data.
  • The contributions of the teacher and the external model are not adequately disentangled. Removing SFT in Table 4 examines the role of initialization during training, whereas removing Prior in Table 2 examines the effect of external API access at inference. These analyses do not isolate the effects of teacher quality from the contribution of Prior.
  • Comparisons with existing methods under matched conditions are insufficient. The paper includes direct-inference baselines using Qwen3 and a reevaluation of IAD-R1 based on Qwen2.5, but it does not compare against existing methods using the same Qwen3 backbone and access to Prior. Such comparisons are particularly important given the substantial performance differences across LLM versions.
  • Some interpretations of performance and generalization remain limited. Without Prior, IndusAgent achieves an average performance of 79.7%, close to the 79.4% obtained by direct inference with Qwen3-VL-Max. The additional accuracy gains over direct inference with the external model therefore appear limited in this setting. Further analysis is needed to identify which interactions between training and tool use account for the full system's 83.5% performance, together with comparisons against baselines given the same access to Qwen3-VL-Max.
Questions for authors

Addressing the following points would support a fairer assessment of the method and its empirical contributions.

  1. Teacher model and data generation: What are the exact model name, version, and generation settings of the teacher used to produce the Indus-CoT trajectories? Could you provide success and failure rates, as well as category-level performance, for trajectories that include actual tool calls and the returned observations?
  2. Contributions of SFT and the teacher: Could ablations with different teacher models help disentangle the sources of the gains from SFT? The substantial decline from 83.6% to 69.5% in Table 4(d) suggests that initialization using teacher-generated data is important. Could you report how IndusAgent's performance changes with teacher quality?
  3. Baselines under matched conditions: Could you compare against IAD-R1 using Qwen3-VL-8B as the backbone and providing the same access to Prior? Do the gains persist relative to a simple policy that combines responses from the same Prior?
  4. Dependence on Prior: How does the method perform when the policy is retrained without using Prior during either training or inference?
  5. Contribution of gating: According to Eq. (6), trajectories with an incorrect final diagnosis receive only RformatR_{format}. If all trajectories in a group are incorrect but satisfy the output-format requirements, their rewards are identical, and all advantages become zero under Sec. 2.4. Does this occur during training? If so, could you report its frequency and examine whether it decreases as training progresses? This analysis would help clarify the contribution of gated GRPO on difficult examples.
6 comments

We appreciate your careful reading and questions about the teacher supervision and training setup.

W1: The teacher model, generation settings, and trajectory quality checks need to be specified.

A1: Thanks for raising the question. We generated trajectories with qwen3.7-max-2026-06-08, leaving generation settings at provider defaults: thinking enabled, temperature 0.6, and top-p 0.95 (sampling defaults; thinking mode). The output-token limit was not explicitly set and used the service default. Request retries used the client's unchanged default behavior. Qwen3-VL-Max serves as the runtime Prior.

The generation teacher selects tools without access to ground-truth annotations. Completed trajectories are screened against source annotations for diagnostic correctness, output format, and required auxiliary fields. Anomalous cases must match the canonical type and achieve box IoU of at least 0.5. Normal cases must end with a normal diagnosis and no final defect box or type. This screen evaluates the final output; intermediate explanations are not scored.

Q1: Which teacher was used and how were trajectories checked?

A2: Thank you for requesting the generation accounting. The generation model and screening criteria are specified above. Initially, 2,747 of 3,000 candidates passed (91.57%). A further 280 replacement attempts yielded 253 accepted trajectories (90.36%), giving 3,000 accepted trajectories from 3,280 attempts (91.46% overall). The final corpus contains 1,500 normal and 1,500 anomalous trajectories. These are candidate-generation counts, including replacements of rejected candidates; they are separate from request retries.

Table 1. Trajectory generation and screening

StageAttemptedAcceptedRejectedAcceptance
Initial normal1,5001,38012092.00%
Initial anomalous1,5001,36713391.13%
Initial total3,0002,74725391.57%
Replacement normal1301201092.31%
Replacement anomalous1501331788.67%
Replacement total2802532790.36%
Combined total3,2803,00028091.46%

W2: It is unclear how the effect of teacher quality is distinguished from the benefit of SFT initialization.

A3: We appreciate this distinction between teacher quality and initialization. The no-SFT ablation tests the value of initializing the policy with teacher-generated trajectories. Removing SFT also removes its supervision of output format and tool use, so these contributions are evaluated together.

Q2: Could the gains come from teacher quality rather than the proposed learning?

A4: Thank you for considering the role of teacher quality. SFT-only versus Full evaluates the subsequent RL stage under the current teacher supervision. The no-SFT ablation evaluates training without the teacher-derived initialization. Together, these comparisons assess the contributions of the training stages with the current teacher data. We have not compared different generation teachers.

W3: Different backbones and access to external information make the baseline comparisons difficult to interpret.

A5: Thanks for pointing out the backbone and information differences. We added common-backbone controls with explicit Prior access. All trained controllers share Qwen3-VL-8B, accepted source images, supervision and update budgets, preprocessing, and evaluation. Prior-enabled variants use the same interface. The IAD-R1-style and AgentIAD-style entries adapt these methods to the common backbone; their objectives or tool configurations can differ. Base plus fixed Prior is evaluated without industrial training.

The trained fixed Crop-plus-Prior controller retains SFT, the full R11 reward, and GRPO. It learns image-dependent crop coordinates and Prior focus, and executes one Crop and one Prior. Each controller receives feedback from its own actions. Full chooses among six tools, while the fixed policy uses two, so this comparison covers both scheduling and available tools.

Table 2. Controller performance and resources

BA is a percentage; means equally weight MVTec-AD, MPDD, VisA, DTD, and SDD. Calls, time, and API charges are per query. Prior is Qwen3-VL-Max; API charges exclude local computation. The supplementary Full result of 83.4 is reported separately from the manuscript result of 83.5.

PolicyMean BALocal / Prior callsSeconds/queryAPI CNY/query
IAD-R1-style without Prior79.20 / 01.070
IAD-R1-style with Prior82.00 / 12.580.0039
AgentIAD-style82.20.65 / 0.301.680.00115
Fixed Crop + Prior82.61 / 12.760.0038
Base + fixed Prior78.80 / 12.470.0037
Full83.40.78 / 0.171.470.0006

Q3: What happens with a common backbone and matched Prior access?

A6: We appreciate the suggestion to compare under matched conditions. Full reaches 83.4 BA, compared with 82.0 for IAD-R1-style with Prior, 82.2 for AgentIAD-style, and 82.6 for the trained fixed policy. Relative to the fixed policy, Full gains 0.8 points while reducing Prior calls from 1.00 to 0.17, latency from 2.76 to 1.47 seconds, and API charges from 0.0038 to 0.0006 CNY per query. These correspond to 46.7% lower latency and 84.2% lower API charges.

Table 3 gives the complete per-benchmark results. The fixed policy is slightly higher on MPDD, at 72.8 versus 72.6; Full has the highest mean BA in this comparison. Measured latency and API charges complement call counts because equal call budgets can still produce different resource use.

Table 3. Per-benchmark balanced accuracy (%)

PolicyMVTecMPDDVisADTDSDDMean
IAD-R1-style without Prior80.570.272.090.782.679.2
IAD-R1-style with Prior82.772.075.894.285.382.0
AgentIAD-style82.972.375.994.585.482.2
Fixed Crop + Prior83.272.876.594.785.882.6
Base + fixed Prior79.567.872.590.583.778.8
Full83.572.676.795.588.783.4

W4: The dependence on Prior and the accuracy-cost comparison need further analysis.

A7: Thank you for focusing on Prior dependence. Retraining without Prior reaches 80.6 BA, compared with 79.8 when Prior is disabled only at test time and 83.4 for Full. Retraining recovers 0.8 points over test-only removal, while Full remains 2.8 points higher. The retrained configuration uses local tools at inference, with zero external Prior calls; teacher supervision and local computation remain part of the system.

Table 4. Prior availability

BA is a percentage; means equally weight MVTec-AD, MPDD, VisA, DTD, and SDD. Calls and time are per query. Prior is Qwen3-VL-Max. Both no-Prior conditions have zero inference API charges, excluding local computation.

ConditionMean BALocal / Prior callsSeconds/query
Full83.40.78 / 0.171.47
Prior disabled only at test79.80.78 / 01.14
Prior removed throughout retraining and test80.60.92 / 01.18

Q4: What if Prior is removed from the entire pipeline?

A8: Thanks for asking about removing Prior throughout the pipeline. We regenerated SFT responses without Prior observations, then removed Prior from RL and testing. The generation teacher and decoding configuration, other tools, source categories, and training budgets remain fixed. This retrained policy reaches 80.6 BA with 0.92 local calls and 1.18 seconds per query.

Its grounding results are 44.4 mean box IoU, 55.7 [email protected], and 70.8 type accuracy. These measure the retrained model's absolute grounding performance. Corresponding Full grounding scores from the same supplementary experiment are unavailable for a paired comparison.

Table 5. Grounding after retraining without Prior

Grounding and type metrics cover all ground-truth-anomalous images, with missing or invalid outputs scored zero. Box IoU is multiplied by 100; [email protected] is the percentage of images with box IoU of at least 0.5. BA, specificity, recall, and type accuracy are also percentages. The macro mean weights the five datasets equally.

DatasetBASpecificityRecallBox IoU[email protected]Type Acc.
MVTec-AD81.873.690.044.256.270.8
MPDD71.449.893.038.546.863.4
VisA71.758.485.040.049.565.1
DTD92.486.298.650.665.480.2
SDD85.778.093.448.760.674.5
Equal macro mean80.669.292.044.455.770.8
Reviewer_KjCjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

IndusAgent trains Qwen3-VL-8B with supervised fine-tuning on 3,000 screened Real-IAD trajectories and two-round tool-augmented GRPO. Its six tools include local image operations and an external Qwen3-VL-Max visual opinion; an accuracy gate conditions auxiliary rewards and tool costs on final diagnostic correctness. Evaluation covers disjoint product categories within a fixed defect vocabulary, reporting 83.5 percent mean balanced accuracy across five benchmarks.

Critical strengths
  1. The source-category exclusion audit and annotation-blind teacher action selection establish a useful transfer setting. Section 2.2 and Appendix B.1 distinguish post-generation screening from annotation-guided trajectory repair.
Critical weaknesses
  1. The evidence does not establish the added value of the proposed inspection design over a simpler, matched agent. AgentIAD already supplies tool-integrated SFT and RL, as Section 4 acknowledges, but is absent from the comparisons. Table 2 instead compares direct models and an IAD-R1 rerun on an older backbone. Tables 4 and 14 remove tool descriptions from a tool-trained checkpoint, jointly changing its prompt and available observations. These interventions establish sensitivity to tool availability, but cannot distinguish learned adaptive inspection from training adaptation, prompt effects, or a fixed crop-and-prior pipeline. A reference-free AgentIAD-style control trained on the same source categories, or a fixed inspection schedule with matched supervision and call budget, would test whether the six-tool policy adds value beyond established machinery.

  2. The main reward novelty is not sufficiently resolved mechanistically. Appendix A explicitly establishes that gating localization and type rewards changes nothing on schema-compliant trajectories: false negatives omit these fields, and normal images receive neither reward. The actual differences concern inconsistent auxiliary outputs and tool bonuses/costs on incorrect trajectories. Table 9 supports the complete objective, but Table 5 removes entire reward terms rather than separating these gating effects. Without rollout frequencies of inconsistent outputs and a control retaining call costs outside the gate, the results cannot establish whether the gain comes from controlling auxiliary credit or changing the cost of unsuccessful exploration. This limits the insight supplied by the central optimization contribution.

Questions for authors

Which teacher model and version generated Indus-CoT?

3 comments

Thank you for your review and suggestions on the evaluation, reward design, and trajectory generation.

K1: Does adaptive inspection improve on a trained fixed policy under a common backbone and matched budgets?

A1: Thank you for highlighting the need for matched controls. Under a common backbone and matched training budgets, Full reaches 83.4 balanced accuracy (BA), compared with 82.6 for the trained fixed Crop-plus-Prior policy. It uses 0.17 versus 1.00 Prior calls per query and takes 1.47 versus 2.76 seconds, reducing latency by 46.7% and API charges by 84.2% (Table 1). This is the highest mean BA among these controls; on MPDD, the fixed policy is slightly higher at 72.8 versus 72.6 (Table 2).

All trained controllers share Qwen3-VL-8B, accepted source images, supervision and update budgets, preprocessing, and evaluation. Prior-enabled variants use the same Qwen3-VL-Max interface. The fixed controller retains SFT, the full R11 reward, and GRPO. It learns image-dependent crop coordinates and Prior focus, and executes one Crop and one Prior for every image. Each controller receives feedback from its own executed actions.

Full selects among six tools, while the fixed policy uses Crop and Prior. The comparison thus evaluates scheduling together with the available tool set. The IAD-R1-style and AgentIAD-style controllers adapt their objectives and tool configurations to the common backbone. Base plus fixed Prior provides a reference without industrial training.

Table 1. Controller performance and resources

BA is a percentage, averaged equally over MVTec-AD, MPDD, VisA, DTD, and SDD. Resource values are per query. Prior is Qwen3-VL-Max; API charges exclude local computation. Supplementary Full (83.4) is reported separately from the manuscript result (83.5).

PolicyMean BALocal / Prior callsSeconds/queryAPI CNY/query
IAD-R1-style without Prior79.20 / 01.070
IAD-R1-style with Prior82.00 / 12.580.0039
AgentIAD-style82.20.65 / 0.301.680.00115
Fixed Crop + Prior82.61 / 12.760.0038
Base + fixed Prior78.80 / 12.470.0037
Full83.40.78 / 0.171.470.0006

Table 2. Per-benchmark balanced accuracy

Scores are percentages; the mean weights the five benchmarks equally. MVTec denotes MVTec-AD.

PolicyMVTecMPDDVisADTDSDDMean
IAD-R1-style without Prior80.570.272.090.782.679.2
IAD-R1-style with Prior82.772.075.894.285.382.0
AgentIAD-style82.972.375.994.585.482.2
Fixed Crop + Prior83.272.876.594.785.882.6
Base + fixed Prior79.567.872.590.583.778.8
Full83.572.676.795.588.783.4

K2: Which reward gates account for the gains, and what happens to group-relative advantages when all sampled answers are incorrect?

A2: We appreciate the question about reward attribution. The added controls favor gating positive auxiliary credit. R10 reaches 83.7 BA, versus 80.6 for the additive reward and 83.4 for R11 (Table 3). R10 also uses fewer calls than R11; the additional cost gate brings no improvement in these experiments.

Let CC denote diagnostic correctness, LL localization reward, TT canonical-type agreement, II the successful-call and positive-margin-change indicator, NN admitted calls, and FF format reward. Define P=0.8L+0.6T+0.15IP=0.8L+0.6T+0.15I and K=0.05NK=0.05N.

The objectives are

R00=C+P−K+F,R_{00}=C+P-K+F,

R10=C+CP−K+F,R_{10}=C+CP-K+F,

R11=C+CP−CK+F.R_{11}=C+CP-CK+F.

R10 versus R00 isolates gating the positive bundle under an unconditional call cost; R11 versus R10 tests the additional cost gate. For schema-compliant incorrect predictions, L=T=0L=T=0: false negatives omit anomaly fields, and truly normal images are ineligible for localization/type rewards. Thus CL=LCL=L and CT=TCT=T on this subset. The experiment tests the positive bundle as a whole, without separately isolating its localization, type, and tool-bonus gates. Higher BA than R00 also accompanies higher FPR, reflecting the recall–false-positive tradeoff.

In an all-incorrect R11 group with equal format rewards, rewards are constant and reward-normalized advantages are zero. For fully compliant outputs, Ci=0C_i=0, Fi=0.05F_i=0.05, and Ri=0.05R_i=0.05. KL regularization can still contribute. Unequal format rewards can produce nonzero advantages; under R10, different call counts can also distinguish incorrect trajectories. We have not measured the frequency of constant-reward groups during training.

Table 3. Reward decomposition

Each objective is trained separately from the same initialization and with the same update budget. BA, FPR, and recall are percentages, with equal means over MVTec-AD, MPDD, VisA, DTD, and SDD. FPR and recall share the averaging hierarchy: BA=(100−FPR+Recall)/2BA=(100-FPR+Recall)/2. Calls are per query; Prior denotes Qwen3-VL-Max.

ObjectiveMean BAFPR / RecallTotal / Prior calls
R00 additive80.616.0 / 77.21.08 / 0.20
R10 unconditional cost83.718.8 / 86.20.85 / 0.14
R11 full83.419.2 / 86.00.95 / 0.17

K3: Which teacher and generation settings were used, and how were the training trajectories filtered?

A3: Thank you for asking for these implementation details. We generated trajectories with qwen3.7-max-2026-06-08, leaving generation settings at provider defaults: thinking enabled, temperature 0.6, and top-p 0.95 (thinking mode; sampling defaults). The output-token limit was not explicitly set, and request retries used the client's unchanged default behavior. Qwen3-VL-Max serves as the runtime Prior.

The teacher selects tools without ground-truth annotations; completed trajectories are then screened against source annotations. Screening checks diagnostic correctness, output format, and required fields. Anomalous cases must match the canonical type and reach box IoU ≥ 0.5. Normal cases must end with a normal diagnosis and no final defect box or type. These criteria assess final outputs; intermediate explanations are not scored by this filter.

Table 4. Trajectory generation and screening

Counts record candidate-generation attempts and accepted trajectories. Replacement attempts fill the shortfall after screening and are counted separately from request retries. The final corpus contains 1,500 normal and 1,500 anomalous trajectories.

StageAttemptedAcceptedRejectedAcceptance
Initial normal1,5001,38012092.00%
Initial anomalous1,5001,36713391.13%
Initial total3,0002,74725391.57%
Replacement normal1301201092.31%
Replacement anomalous1501331788.67%
Replacement total2802532790.36%
Combined total3,2803,00028091.46%

Thank you for the additional comparisons and reward ablations. They clarify the practical profile of the system, particularly its efficiency in tool and Prior usage. However, my concern about how the contribution is framed remains. The new controls suggest that the main benefit of adaptive inspection lies in efficiency rather than in a qualitatively stronger diagnostic policy. The reward analysis also indicates that the simpler gating variant is sufficient, so the proposed cost gate is not essential. Taken together, the work reads more as a well-engineered combination of existing agentic components than as a methodological advance. Its core argument would need to be reframed around what is genuinely new. I therefore maintain my original recommendation.

Comments on the paper

1 comment
Area_Chair_iLJxOpenReview ↗

Please engage with the authors rebuttal by Nov 18

Dear Reviewers,

The authors have posted their rebuttal. Please engage with it by November 18, 2026, noting which of your concerns are resolved or which remain. Even brief engagement helps greatly.

Best regards,
Area Chair

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

Related papers

Discussion (0)

Sign in to comment.