acceptodds
Under review as a conference paper at ICLR 2027

APPLENMR-TEXT: A COST-SENSITIVE LF-NMRBENCHMARK FOR APPLE MOLDY-CORE SCREENING

Abstract

Apple moldy-core screening is a small-data, cost-sensitive task: the defect is in-ternal, destructive verifcation is needed for labels, and expert diagnostic notes areunavailable as inputs for new fruit. We introduce AppleNMR-Text, a benchmarkof 237 Red Fuji apples with LF-NMR slices, destructively verifed labels, andexpert-reviewed text, together with fxed cross-validation and quality-gated utilityevaluation. As a documented image-input benchmark model, we study a two-stage pipeline: dual-memory retrieval guides structured rationale generation, thena classifer combines the image, generated rationale, and sanitized evidence. Onthe reported fve-fold evaluation, this model obtains TAAPM 393.733± 13.973,versus 158.40± 224.01 for vision-only VGG19 and 155.33± 61.98 for directlyfne-tuned Qwen3-VL-8B. Dual-memory retrieval outperforms either single memory on TAAPM. Balanced accuracy is 0.700± 0.019, so the utility advantageshould not be read as uniform superiority on conventional metrics. AppleNMR-Text provides a controlled testbed for image-only screening and for separating anexpert-text oracle from a deployable input protocol; its single-season cohort doesnot establish cross-domain generalization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Apple moldy core disease, LF-NMR imaging, Provenance aware dual memory RAG, Rationale then classify

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

The paper focuses on the real scenario of apple mold core, which has internal defects, destructive labels, small samples, and cost sensitivity. A dedicated evaluation set was constructed, including LF-NMR images, expert text, fixed partitioning, and TAAPM protocol. The paper also conducted a systematic internal comparison and ablation under a unified setting, covering multiple baselines, retrieval sources, rational forms, and training objectives, with clear domain resource value. However, the scope of application of this work is relatively narrow, covering only a single variety, modality, season, and collection conditions, and the impact of general ML is limited; There are only 237 apples in the queue, with only 15-16 diseased samples per fold. As a core indicator, TAAPM still relies on a set of unproven fixed contract parameters and fixed thresholds, lacking sensitivity analysis. Therefore, the current results can only support the cost sensitive utility under this single queue and specific contract settings, and are not sufficient to support the claims of a universal benchmark or deployable method.

Critical strengths
  1. The paper focuses on the internal defect screening task of apple mold core. This task has several clear and genuine characteristics: defects are located inside the fruit, and there may be no significant changes on the surface; The label needs to be scanned and obtained through destructive inspection; The number of diseased samples is small and the categories are unbalanced; The cost of missed and false detections is asymmetric; Unable to obtain expert text during testing.
  2. The paper compared multiple types of baselines under a unified evaluation protocol, including zero shot and fine tuned versions of vision only VGG19 and Qwen3-VL 2B/4B/8B GPT-5-Instant、GPT-5-Thinking,And the author's two-stage system. The paper also includes text only BERT, Stage II oracle, and human expert controls, but these controls use additional text input and cannot be directly compared equivalently with the image only model.
Critical weaknesses
  1. The algorithm components of the paper are mainly a combination of existing methods, including Qwen3-VL, VGG19, BERT, LoRA, contrastive learning, EMA, and gated fusion; The paper did not propose new basic models, network structures, or learning objectives. Its contribution is more suitable for positioning as domain datasets, evaluation protocols, and system baselines, rather than innovation in general ML methods.
  2. The data only covers Red Fuji apples, mold core binary classification, LF-NMR modality, as well as single season, refrigeration protocol, operator, and collection day. The paper did not validate other apple varieties, other internal defects, other fruits, other scanners, or real production line conditions. Therefore, the external applicability of its conclusion is limited.
  3. There are only 237 apples in the queue, of which 79 are diseased samples. Each test set consists of 47-48 apples, with only 15-16 diseased samples. The statistical power is limited, and the differences in model ranking and small sample conditions may be unstable.
  4. The paper also did not provide sensitivity analysis of TAAPM to parameter perturbations, threshold scanning, calibration experiments under different disease rates, or comparison with accuracy/F1 ranking systems.
Questions for authors

Please refer to 'Weaknesses'.

1 comment
AuthorsOpenReview ↗

Response to Reviewer BiKM

We sincerely thank you for your careful review and constructive comments on the contribution and evaluation. We apologize for our delayed response.

  1. Contribution (Weakness 1).

    Our contribution combines an LF-NMR dataset and benchmark, an evaluation protocol, and a documented reference system with component analysis. It links images, destructive labels, and expert-reviewed text; separates image-input and expert-text evaluations; and compares components under a shared quality gate.

    Revision: Abstract and Sections 1–2.

  2. Evaluation cohort and statistics (Weakness 3).

    Each reported model metric covers all 237 apples per seed, including 79 diseased apples. The 47–48 apples belong to one fold. Metrics are computed after pooling five folds; mean and SD summarize three seed-level results:

    ItemProtocol
    Independent cohort237 apples: 79 diseased and 158 healthy.
    Cross-validationFive apple-level folds within each seed.
    Individual test fold47–48 apples, including 15–16 diseased apples.
    Evaluation per seedPool all five held-out folds: 237 predictions, one per apple.
    Metric calculationCompute Bal-Acc, F1, MCC, ACC, and TAAPM once on the pooled predictions. Apply the TAAPM quality gate after pooling.
    Repeated model runsThree random seeds: 2024, 2025, and 2026.
    Reported mean and SDCalculated across the three seed-level metric values, rather than across fold-level values.

    Revision: Section 3/Table 1; Tables 3–5 captions; Supplementary E.1 and opening of F.

  3. Controlled cohort (Weakness 2).

    Standardized cultivar, season, batch, storage, operator, and acquisition reduce confounds. This enables controlled method comparison. Conclusions apply to this cohort and setting; external applicability needs separate evaluation.

    Revision: Sections 3 and 7; Supplementary F.1.

  4. TAAPM and conventional rankings (Weakness 4).

    TAAPM combines shipment quality and procurement efficiency. Missed defects can violate the quality cap; rejecting healthy fruit raises procurement costs. Its operational objective can rank models differently from accuracy or F1. Table 3 reports both. Utility can be negative after the quality gate is passed if procurement costs exceed modeled revenue.

    Revision: Sections 5 and 6.1/Table 3; Supplementary E.1–E.3.

  5. Operational parameters (Weakness 4).

    All methods use prespecified parameters. The 10% operational prior is the three-year factory average after upstream screening; the 5% shipment cap is the buyer's requirement. This prior differs from 79/237 in the cohort; utility refers to this factory setting.

    Revision: Sections 5 and 7; Supplementary E.1–E.3.

  6. Threshold sensitivity (Weakness 4).

    We test 0.48, 0.50, and 0.52; the main threshold remains prespecified at 0.50. Four baselines change; other configurations remain unchanged. The full system reports TAAPM 393.733 ± 13.973 throughout this range. The hard gate can amplify small classification changes.

    Supplementary Table S4. Threshold sensitivity. Mean ± SD across three seeds; 237 pooled apples per seed.

    ModelThresholdBal-AccF1MCCACCTAAPM
    VGG190.480.705 ± 0.0440.617 ± 0.0510.390 ± 0.0840.681 ± 0.039196.457 ± 194.710
    VGG190.500.703 ± 0.0340.604 ± 0.0470.389 ± 0.0530.681 ± 0.035158.40 ± 224.01
    VGG190.520.694 ± 0.0380.606 ± 0.0440.369 ± 0.0720.677 ± 0.044201.931 ± 198.020
    Qwen3-VL-2B (FT)0.480.595 ± 0.0140.433 ± 0.0200.205 ± 0.0300.665 ± 0.0140.000 ± 0.000
    Qwen3-VL-2B (FT)0.500.594 ± 0.0110.431 ± 0.0210.202 ± 0.0230.665 ± 0.011-45.875 ± 79.443
    Qwen3-VL-2B (FT)0.520.592 ± 0.0110.427 ± 0.0140.199 ± 0.0260.664 ± 0.0130.000 ± 0.000
    Qwen3-VL-4B (FT)0.480.571 ± 0.0270.405 ± 0.0660.147 ± 0.0440.633 ± 0.0080.000 ± 0.000
    Qwen3-VL-4B (FT)0.500.568 ± 0.0200.404 ± 0.0460.142 ± 0.0360.632 ± 0.0040.000 ± 0.000
    Qwen3-VL-4B (FT)0.520.572 ± 0.0230.403 ± 0.0640.151 ± 0.0360.637 ± 0.0110.000 ± 0.000
    Qwen3-VL-8B (FT)0.480.633 ± 0.0240.468 ± 0.0530.315 ± 0.0410.719 ± 0.0130.000 ± 0.000
    Qwen3-VL-8B (FT)0.500.629 ± 0.0190.478 ± 0.0320.284 ± 0.0400.718 ± 0.012155.33 ± 61.98
    Qwen3-VL-8B (FT)0.520.628 ± 0.0210.454 ± 0.0490.311 ± 0.0300.719 ± 0.0090.000 ± 0.000

    Revision: Section 5; Supplementary E.3/Table S4.

Thank you again for your thoughtful feedback. We welcome further questions and suggestions and look forward to continuing the discussion.

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

The paper introduces a dataset and a model for detecting moldy core in Fuji apples from LF-NMR images, without cutting the apples open. The dataset contains 237 apples from a single batch. It includes LF-NMR images, ground-truth labels obtained by cutting the apples open, and expert-reviewed text descriptions. The model is optimized for a TAAPM score, which appears to reflect buyer requirements: enough good apples while keeping moldy apples below a level α. It is compared against baseline models and shows an improved TAAPM score but no improvement on the other metrics.

Critical strengths

The proposed method seems to improve on the task-specific TAAPM score.

Critical weaknesses

• Dataset size and significance: The dataset is too small and not varied enough: all apples are the same type and come from the same batch, the same storage, and the same year. It would need to be expanded to be a more useful general resource for the community, which should not be expensive in this domain. • Overengineered model: The model is complex and hard to follow from the text. There is limited evidence that it is more useful beyond optimizing TAAPM. It does not seem to outperform the baseline methods on the other reported metrics. • Clarity: Many acronyms (e.g., LF-NMR, TAAPM) are not properly introduced, and the text is hard to follow. TAAPM in particular is only described fairly late in the text. Without improvements to the clarity of the text, such as a more thorough description of the method and its motivation, it is hard to tell what the model does and why it is necessary.

Questions for authors

Clarifications: • What are LF-NMR images, beyond the statement that they expose internal structure? Please properly introduce them more thoroughly. • How are the text descriptions processed? Please clarify lines 82–84: what is the controlled vocabulary, and how are potential confounders and uncertainty accounted for? • Why are this strategy and imaging modality preferable to those used in previous work? Please put the benefits of this work more concretely in perspective.

3 comments
AuthorsOpenReview ↗

Response to Reviewer GWhy

We sincerely thank you for your thoughtful review and constructive suggestions on the benchmark's value, the model's purpose, and presentation. We apologize for our delayed response.

  1. Dataset significance and collection cost. AppleNMR-Text provides a dedicated LF-NMR benchmark linking internal images, destructively verified disease labels, and expert-reviewed diagnostic text. The controlled cultivar, batch, harvest year, storage, and acquisition setting isolate methodological comparisons. A fully annotated apple costs approximately USD 12, including scanning, destructive verification, and expert review. Expanding the resource requires repeating this complete process. The current cohort supports controlled comparisons and provides a common starting point for further LF-NMR screening research. Revision: The Abstract and Sections 1 and 3 clarify the resource and protocol contributions; Supplementary Section F.1 explains the controlled design and collection cost.

  2. Model purpose, complexity, and usefulness. Stage I retrieves diagnostic evidence and generates an inspectable rationale; Stage II integrates the image and generated text for the screening decision. The objective is decision utility under a shared quality constraint. Table 3 reports F1 0.608 ± 0.022 for our model versus 0.604 ± 0.047 for VGG19, alongside the conventional metrics and TAAPM. Table 5 compares retrieval sources and rationale inputs: the full dual-memory configuration reports higher TAAPM than either single memory. These comparisons give each major component a testable role in the benchmark. Revision: Sections 4 and 6.1–6.3, Figures 2 and 4, and Tables 3–5 clarify the information flow and provide the corresponding component evidence.

  3. What LF-NMR images measure and their relation to prior work. Low-field nuclear magnetic resonance detects hydrogen-nucleus signals, with fruit-image contrast largely associated with water protons. Our T₂-weighted images provide spatial information about water distribution and tissue microenvironments. Destructive inspection supplies the final disease label. The benefit is a spatial internal imaging representation paired with verified pathology and expert interpretation. This complements the Vis/NIR spectra and vibration responses used in prior work; the present study evaluates LF-NMR within this benchmark. Revision: Sections 1–3 introduce the modality and its relation to prior studies; Supplementary Section F.1 explains acquisition and the physical interpretation.

  4. Controlled vocabulary, confounders, uncertainty, and text processing. Expert records standardize signal-pattern descriptors, including distribution, heterogeneity, attenuation, and asymmetry, together with spatial tags. Confounder tags record storage, acquisition, or physiological factors. Uncertainty notes preserve ambiguous or conflicting evidence. Experts normalize synonyms and adjudicate unsupported or inconsistent statements. Findings, supported claims, and uncertainty are retained in the compact Stage II text input.

    A held-out apple's own expert record and label are excluded from standard image-input inference. Retrieved cases are separately sanitized to remove verdict-equivalent information. Revision: Section 3 and Sections 4.1–4.2 clarify the text roles; Supplementary Sections A.3–A.5, B.4, C.3, and D.2 provide the schema, cleanup rules, and annotation protocol.

  5. Acronyms and presentation. The Abstract expands both low-field nuclear magnetic resonance (LF-NMR) and the Task-Aware Apple Pathology Metric (TAAPM). The Introduction connects the sensing task to shipment-quality requirements, Figure 2 summarizes the two stages, and Section 5 gives the utility calculation. Detailed procedures remain in the supplement. Revision: The Abstract, Sections 1, 4, and 5, and Figure 2 clarify terminology, motivation, and the method's information flow.

Thank you again for helping us clarify the benchmark's motivation and presentation. We welcome further feedback and look forward to continuing the discussion.

Reviewer_GWhyOpenReview ↗

Can't see the updated text

Hi, I'm trying to check the updates but can't find the difference in the updated text. For example, I can't see the changes to the abstract described above. Have the text been updated?

AuthorsOpenReview ↗

Follow-up Response to Reviewer GWhy

Thank you for the time and care you have put into reviewing our work. We appreciate the effort that a careful review requires, and we are grateful that you took another look at the manuscript and followed up with us.

1. Questions you raised

  • Was the abstract updated? Yes. In the revised PDF, p. 1, lines 15–19 spell out low-field nuclear magnetic resonance (LF-NMR) and Task-Aware Apple Pathology Metric (TAAPM).
  • Where are the imaging and text clarifications? Sections 1–2 and Appendix F.1 (p. 30) explain what LF-NMR measures, how it complements prior sensing approaches, and why we used controlled collection conditions. Appendix D.2 (p. 25) explains the vocabulary, confounders, and treatment of uncertainty.
  • What do the two stages contribute? Figure 2 (p. 4) shows how retrieved evidence informs rationale generation and classification. Table 5 (p. 9) reports higher utility when both retrieval sources are used together under the same evaluation conditions.

2. Possible follow-up questions

  • What can other teams do with the dataset? The released images and verified labels, together with the documented splits and protocol, let teams reproducibly compare new screening methods on the same 237 apples without repeating scans or destructive inspection.
  • How can the utility comparison be checked? Section 5 (p. 6) gives the formula and shared contract parameters. Appendix E.1 (p. 27) explains how we pool the five held-out folds within each seed. Together, they specify how to recalculate the TAAPM values in Table 3 from the held-out predictions at the fixed threshold of 0.5.

Thank you again for the constructive discussion. We would be glad to continue the conversation, answer any further questions, and learn from your feedback.

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

This paper introduces the AppleNMR-Text dataset, which contains low-frequency NMR images of 237 apples, destructive-tested labels, and expert-reviewed text. The paper designs a two-stage process: first, interpretation is generated using retrieved evidence, followed by classification; the inference stage requires only images as external input. Five-fold cross-validation with three random seeds shows improved performance of TAAPM under the given protocol, but traditional classification metrics do not consistently outperform the baseline model.

Critical strengths

The dataset pairs imaging with destructively verified labels and expert text, providing a useful resource for internal-defect screening. The distinction between image-input systems and expert-text reference models makes the evaluation practically relevant. The baseline comparisons and ablations examine important components, including rationale inputs and retrieval sources. The authors also clearly acknowledge the limits of the utility gains and the single-cohort evaluation.

Critical weaknesses

The second phase uses TAAPM on the validation set to select checkpoints, but it's unclear whether VGG19 and the fine-tuned Qwen3-VL also use the same criteria. Some retrieval and text processing settings are not clearly explained, nor is it clear whether TAAPM is calculated individually for each fold or based on the aggregated prediction results. This metric includes a non-linear quality gate, which limits independent validation of the results, so different algorithms may yield different results.

Questions for authors

Do the VGG19 and fine-tuned Qwen3-VL baseline models also use validation set TAAPM to select checkpoints? If not, does the paper's reported advantage still hold when the selection criteria are consistent? Besides using randomized permutation inference for ablation, could the authors analyze situations where generative inference errors occur? Can the final classifier use image evidence to correct these errors, or is it more likely to follow erroneous inferences?

1 comment
AuthorsOpenReview ↗

Response to Reviewer WVhc

We sincerely thank you for your careful review and constructive questions on comparison fairness, metric aggregation, and incorrect rationales. We apologize for our delayed response.

  1. Checkpoint selection and the fixed decision threshold. All trained models use the same 50-epoch training budget and epoch-50 checkpoints. The decision threshold is fixed at τ=0.5 across all main model comparisons. Neither the checkpoint nor the threshold is chosen from validation or test performance, and no early stopping is used. The supplementary threshold analysis does not select or replace this operating point.

    The earlier validation-TAAPM checkpoint statement was retained from a superseded protocol. We apologize for this inconsistency. Under the stated fixed protocol, Table 3 reports TAAPM 393.733 ± 13.973 for our model, 158.40 ± 224.01 for VGG19, and 155.33 ± 61.98 for Qwen3-VL-8B (FT). Revision: Sections 4.2 and 5 (fixed threshold); Supplementary F.2–F.3 (final-epoch checkpoints) and E.3 (additional threshold analysis).

  2. Statistical interpretation and TAAPM aggregation. Each model metric uses all 237 pooled held-out predictions per seed. The 47–48 apples belong to one fold. TAAPM and its quality gate are computed after pooling; mean and SD are then calculated across three seed-level results. Evaluation and statistical reporting protocol

    ItemProtocol
    Independent cohort237 apples: 79 diseased and 158 healthy.
    Cross-validationFive apple-level folds within each seed.
    Individual test fold47–48 apples, including 15–16 diseased apples.
    Evaluation per seedPool all five held-out folds: 237 predictions, one per apple.
    Metric calculationCompute Bal-Acc, F1, MCC, ACC, and TAAPM once on the pooled predictions. Apply the TAAPM quality gate after pooling.
    Repeated model runsThree random seeds: 2024, 2025, and 2026.
    Reported mean and SDCalculated across the three seed-level metric values, rather than across fold-level values.

    Revision: Section 3; Tables 3–5 captions; Supplementary E.1 and opening of F (pooling and reporting order).

  3. Effect of expert-verified incorrect rationales. We increased the Stage I sampling temperature and had domain experts verify substantive errors in the resulting rationales. We then supplied them to the corresponding frozen Stage II classifiers, keeping images and fold-specific Evidence Packs unchanged. All conditions cover the same 237 apples.

    Table S5 reports lower balanced accuracy, F1, and MCC under incorrect rationales, while accuracy increases slightly. The controlled intervention shows that incorrect intermediate text can degrade the final multimodal decision. Images and retrieved evidence remain available in this test; VGG19 is an independent vision-only reference.

    Supplementary Table S5. Robustness to incorrect generated rationales.

    ConditionBal-AccF1MCCACCTAAPM
    G1: Original rationale0.700 ± 0.0190.608 ± 0.0220.382 ± 0.0380.700 ± 0.019393.733 ± 13.973
    G2: Shuffled rationale0.689 ± 0.0200.596 ± 0.0230.360 ± 0.0400.690 ± 0.021260.266 ± 225.585
    G3: Incorrect rationale0.666 ± 0.0240.549 ± 0.0360.322 ± 0.0410.706 ± 0.012213.067 ± 79.903
    Vision-only (VGG19)0.703 ± 0.0340.604 ± 0.0470.389 ± 0.0530.681 ± 0.035158.40 ± 224.01

    Mean ± SD across three seeds; 237 pooled apples per seed. VGG19 is an independent reference. Revision: Section 6.3; Supplementary G/Table S5 (full-cohort comparison of original, shuffled, and incorrect rationales).

  4. Retrieval settings and text processing. The revised supplement specifies a case candidate pool of 20, a cosine-similarity diversity threshold of 0.95, and five retained knowledge items plus five cases. The Stage I Evidence Pack budget is 2048 generator-tokenizer tokens. The combined Stage II text input is capped at 512 BERT tokens, including special tokens.

    Held-out apples are excluded from retrieval, and training queries exclude all images sharing the query's apple ID. Case verdicts are removed. Invalid citation identifiers are deleted before compression; claims left without valid support are removed, and the cleaned output is checked against the acceptance conditions. Provenance identifiers remain audit metadata. Revision: Section 4.1; Supplementary A.5, B.3–B.4, C.3, and F.2–F.3.

Thank you again for helping us clarify the evaluation protocol. We welcome further questions and look forward to continuing the discussion.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General comment to the SAC and reviewers

We sincerely thank the SAC and all reviewers for the time and care devoted to evaluating our work, and for their thoughtful questions and constructive feedback. We apologize for the delay in posting our responses and appreciate your patience.

The central contribution is a new benchmark connecting verified internal pathology, expert diagnostic knowledge, and screening decisions under explicit quality constraints. AppleNMR-Text couples LF-NMR images with destructive labels and expert-reviewed text, specifies apple-level evaluation with image-input inference, and provides a documented reference pipeline with systematic ablations. Its scientific value is to enable controlled, comparable study of how retrieved evidence and generated rationales affect operational decision utility. The contribution combines a new data resource, an evaluation protocol, and evidence about system components.

We clarify three shared questions:

  1. Controlled data and practical inputs. The standardized cohort isolates methodological comparisons. Each fully annotated apple costs approximately USD 12 for imaging, destructive verification, and expert review. A held-out apple contributes only its image; its own expert note and label are excluded from inference. Revision: Sections 1 and 3; Supplementary Sections A and F.1.

  2. A unified evaluation protocol. Each seed evaluates all 237 apples through pooled out-of-fold predictions. Metrics are computed once per seed, then summarized across three seeds. All trained models use the same 50-epoch budget and epoch-50 checkpoints. The decision threshold is fixed at 0.5 for all main model comparisons. Neither checkpoints nor thresholds are selected using validation or test labels. Revision: Sections 3–5; Supplementary Sections E.1 and F.2–F.3.

  3. The role of TAAPM. TAAPM evaluates screening under shared shipment-quality and procurement-cost requirements. Its parameters are fixed from factory records and buyer requirements; conventional metrics are reported alongside it. Revision: Section 5 and Section 6.1/Table 3; Supplementary Sections E.1–E.2.

This discussion round adds two experiments and tabulates their results:

  1. Expert-verified incorrect rationales. The full-cohort intervention keeps images, retrieved evidence, and Stage II classifiers fixed. Reported F1 decreases from 0.608 ± 0.022 to 0.549 ± 0.036, showing that intermediate text quality affects final decisions. Revision: Section 6.3; Supplementary Section G/Table S5.

  2. Local decision-threshold sensitivity. We evaluate 0.48, 0.50, and 0.52. Table S4 lists the four baseline configurations whose results change. The full system retains its reported TAAPM of 393.733 ± 13.973 throughout this range. Revision: Section 5; Supplementary Section E.3/Table S4.

Detailed point-by-point responses are provided in the comments below. We welcome further questions and feedback and look forward to continuing the discussion.

AuthorsOpenReview ↗

AppleNMR-Text: open benchmark contribution and further discussion

Dear SAC, AC, and Reviewers,

Thank you again for your time and care in evaluating our submission.

To our knowledge, AppleNMR-Text is the first openly available LF-NMR dataset for apple moldy-core disease screening. It pairs images from 237 apples with destructively verified labels and expert-reviewed diagnostic text. Together with documented baselines and a shared evaluation protocol, it provides a reusable testbed for studying how expert knowledge supports image-input screening under explicit shipment-quality and cost requirements.

Building this resource costs approximately USD 12 per fully annotated apple—about USD 2,800 for the cohort, covering imaging, destructive verification, and expert review. Sharing these data allows other researchers to investigate this task without repeating the full acquisition and annotation process. We will continue maintaining the dataset to support reuse and reproducible comparisons.

On 10 November, we posted point-by-point responses to all three reviews, clarified the evaluation and implementation details, and reported two additional experiments: expert-verified incorrect-rationale robustness and local decision-threshold sensitivity.

With three days remaining for author responses, we would greatly appreciate feedback on any remaining concerns. Receiving feedback while time remains would help us prepare further clarifications and assess any additional experimental checks.

Thank you to the reviewers for their consideration and to the AC and SAC for overseeing the discussion. We welcome follow-up questions and look forward to further exchange.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.