acceptodds
Under review as a conference paper at ICLR 2027

SABRE: Scalable and Automated Benchmarking of VLMs under Stress

Abstract

Vision-language models (VLMs) advance rapidly, but fixed benchmarks can saturate and constructing targeted stress tests remains labor-intensive. Collecting and annotating existing images does not ensure that they satisfy task-specific visual conditions or challenge current models. To solve this, we present SABRE, a scalable, semi-automated pipeline that converts Test Primers (Markdown Task Designs with Data Schemas) into structured specifications, generated or edited images, and question–answer pairs. A Filtering VLM screens candidates for obvious generation defects and prunes readily solved cases, focusing human review on challenging candidates. Reviewers verify task validity, correct annotations, and perform localized image repair when needed. As our main case study, we instantiate SABRE-Prior to test whether VLMs follow visual evidence over world priors—learned expectations about familiar objects and scenes. Its 600 images and 1,000 questions span four subsets: Context, Texture, Attribute, and Language Elicitation. Across six VLMs, macro-average accuracy ranges from 17.8% to 31.3% (22.6% mean). Pilot instantiations on Counting and Spatial reasoning, aided by procedural scene extensions, further demonstrate how the shared workflow generalizes to settings requiring precise quantitative and geometric control. Together, these components provide a reusable framework for constructing and refreshing targeted VLM stress tests as model capabilities evolve.

Primary area:
datasets and benchmarks
Keywords:
Vison Language Model, Benchmark, Visual Evidence, Stress test

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_546UOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

SABRE constructs VLM stress tests from task specifications using image generation/editing, model-based difficulty filtering, human verification, and localized repair. The benchmark contains 600 images and 1,000 questions across context, texture, attribute, and language-elicitation subsets. Six SOTA VLMs (Claude, Kimi, Qwen, Gemini, GPT, Grok) are evaluated. Additionally it contains more difficult stress-test instantiations with 20 samples for counting and spatial where the models answer at most 1/20 counting samples and 0/20 spatial samples correctly. Comparison against baselines such as HallusionBench and VLMBias is done.

Critical strengths
  • Produces challenging visual tests with human verification. Also includes comparisons with existing benchmarks, alternate filtering models, and a repair-preference study. The counting and spatial extensions also show that the workflow can support different tasks.
Critical weaknesses
  • Contribution of the data pipeline is not sufficiently isolated where it doesnt establish how much filtering and repair improve benchmark quality.
  • Additional counting and spatial tests contain only 20 samples each and produce almost uniformly zero accuracy. Without independent human performance on these samples (and a more detailed analysis of model errors), it is hard to tell what these results mean. Generally, the paper provides not enough quantitative analysis of where and why models fail.
Questions for authors
  • How accurate do human evaluators answer the 20-sample counting and spatial tests?
  • What is the total construction cost and measured human time per accepted sample?
8 comments

(1) How do filtering and repair contribute to benchmark quality?

Our appendix reports quantitative evidence for both stages. In Context, difficulty filtering reduced the number of candidates requiring human review by 33.2%. About 21 of the 100 retained edited images underwent localized repair. In a separate blind study, 374 of 400 selections (93.5%) favored SABRE's repair results.

Difficulty filtering. A stress-test example needs to satisfy two criteria. It must support a clear question and a correct answer, and it must expose a model failure. Even a fully valid image can be easy for a VLM. SABRE screens for model failures before human review, so reviewers can focus on candidates that are more likely to meet the difficulty criterion. In Context, this reduced the pool requiring manual review from 220 to 147 cases (Appendix D.4). Human reviewers then checked the remaining images, questions, and reference answers. The 33.2% figure measures the reduction in the review pool. The independent human evaluation in Appendix E.1 and Table 9 provides evidence that the final SABRE-Prior examples are solvable by humans.

Localized repair. Image generators can miss precise visual requirements. For example, a prompt asking for a zebra with ten legs may produce the wrong number of visible legs. If the defect is local, reviewers can mark and repair the affected region, preserve the rest of the image, and verify the result before accepting it. About 21 retained Context images followed this repair-and-recheck process.

Repair quality. Figure 11 shows the instructions and interface used in the blind study. Participants saw an image with an unwanted object marked and selected the result that best removed the object while preserving unrelated content and looking natural. The four methods were shown anonymously and in randomized order. The hard-paste and soft-mask methods used the same generated patch, allowing participants to compare the compositing methods directly. Twenty participants each evaluated 20 cases, and 374 of 400 selections (93.5%) favored SABRE's result (Appendix D.3, Table 6).

Figures 9 and 10 show how the review platform supports this process. Reviewers can inspect the image, question, reference answer, and model prediction, correct annotations, and review a localized repair in the same interface.

(2) How well do humans perform on Counting and Spatial, and what do the near-zero model scores tell us?

Independent human evaluators achieved 90% on Counting and 100% on Spatial on the two 20-example pilots. The two evaluators had no role in constructing the data and saw only the images and questions, following the protocol used for SABRE-Prior in Appendix E.1. We will add these scores and the evaluation procedure to the revision.

The two pilots show that SABRE's generation, screening, verification, and repair workflow can support tasks beyond reliance on world priors. Appendix F also reports larger procedural versions that vary scene complexity. The programs render scenes and compute reference answers from the scene specifications without paid image-generation API calls.

Procedural taskDataset sizeComplexity levelsQwen 3.5 27B accuracy at those levels
Counting (Table 12)200 images, 200 questions20, 50, 100, 150 objects36%, 0%, 0%, 0%
Spatial (Table 13)200 images, 400 questions2×2×2, 4×4×4, 7×7×7, 10×10×10 grids65%, 24%, 3%, 1%

The procedural scenes and questions were fixed before model evaluation. These results show where performance drops as scenes become denser or grids become larger. We will point to these results more clearly in the main text.

(3) What are the construction costs and human time per accepted example?

For Context, Appendix D.4 and Table 7 report approximately $0.44 in initial image-output costs per accepted case. The appendix also reports typical times for individual review and repair operations.

Context construction metricReported value
Initial image outputs440, including a base image and an initial edited image for each of 220 candidates
Estimated cost of initial image outputs$44.44
Estimated initial image-output cost per accepted case$0.44, across 100 accepted cases
Inspect a candidate and record an initial decisionTypically under 5 seconds
Mark a region and enter a repair instructionTypically under 15 seconds
Wait for the repair API to return a resultAbout 5 seconds under normal network conditions

Image generation and editing are the main paid costs in our workflow. At fixed model and output settings, the cost of initial image generation scales with the number of outputs. The reported subtotal includes initial outputs for candidates that were later discarded. It covers initial image outputs only.

The review and repair durations are experience-based estimates for individual operations. Total active human time per accepted case was not systematically measured. We will make these distinctions clearer in the revision.

Reviewer_7jWLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SABRE, a modular, semi-automated pipeline for constructing targeted stress tests for vision-language models. It integrates task specifications, image generation/editing, model-based quality and difficulty filtering, and human verification and localized repair. Its main benchmark, SABRE-Prior, contains 600 images and 1,000 questions across four subsets testing whether models follow visual evidence over learned priors and language-induced assumptions. Six evaluated VLMs achieve macro-average accuracies of 17.8%–31.3%. Independent human evaluation and a repair user study assess benchmark validity and repair quality, while Counting and Spatial extensions explore the workflow’s applicability to additional tasks.

Critical strengths
  1. The unified tool integrates generation, review, and localized repair, offering practical value and demonstrating the feasibility of constructing stress tests across multiple tasks.
  2. Independent human evaluation supports the annotation validity and human solvability of the main benchmark. Models not used for filtering also perform poorly, supporting the cross-model difficulty of the examples. The repair user study provides additional evidence of localized repair quality.
Critical weaknesses
  1. The incremental value and efficiency gains over existing pipelines remain insufficiently validated. AutoBench-V and AutoHallusion have explored related automated benchmark-construction pipelines, yet the paper neither cites these works nor adequately discusses how SABRE differs from them. The reduction in candidates requiring review, experience-based time estimates, and partial image-generation costs reported in Appendix D.4 do not substitute for a comparison of end-to-end construction costs per example. In particular, the paper lacks a baseline combining an existing pipeline with the same hard-example filtering procedure. Such a baseline would help isolate additional benefits of SABRE. For example, whether it improves construction efficiency for examples meeting equivalent requirements for validity, annotation quality, and difficulty.
  2. The interpretation of the overall score is limited by filtering and metric design. The authors acknowledge filtering dependence and differences in metrics across benchmarks, so the score gaps in Table 2 cannot directly quantify an increase in difficulty. Furthermore, a strategy that completely ignores the images, always selects “unknown” for multiple-choice questions, always answers “yes” for yes/no questions, and returns a constant for counting questions would achieve a macro-average score of at least 25% under the scoring rules, exceeding five of the six evaluated models. The current Language subset and its contribution to the macro-average cannot distinguish evidence-based recognition of insufficient information from unconditional abstention without using the image.
  3. The automated filtering description needs greater precision. Figure 2 and Appendix B suggest case-level screening using All4, whereas Section 3.4 defines retention at the individual-question level. These descriptions should be reconciled by explicitly specifying whether a complete paired case, including correctly answered probes, is retained whenever at least one probe fails. The prompts and pass/fail criteria for the preliminary automated image-quality check are also insufficiently specified, limiting reproducibility of this stage.
Questions for authors
  1. On a shared task, with the main foundation models, hard-example filtering rules, and validity and annotation-quality requirements matched, how does SABRE compare with the closest existing construction pipeline in terms of total model-call cost and human labor per valid challenging example, as well as the yield of such examples?
  2. Could the authors report the image-blind baseline and test whether the Language results and overall conclusions hold with matched answerable controls?
8 comments

(1) What SABRE adds to existing pipelines and how it affects construction efficiency

We thank the reviewer for pointing out AutoBench-V and AutoHallusion and asking how SABRE compares with existing workflows when the goal is to produce valid, challenging examples. We will cite both works and clarify the differences in the revision.

What SABRE adds. AutoBench-V generates evaluations on demand for a requested capability. It generates images at different difficulty levels, checks image–description alignment using VQA, and constructs questions and answers. AutoHallusion uses targeted image edits, such as object insertion and removal, to create conflicts between visual evidence and learned priors. Its use of model failures to select hallucination cases is closely related to SABRE-Prior.

SABRE is a reusable framework for building targeted stress tests. It combines task specifications, model-based difficulty screening, human verification, and localized repair. The review platform lets users inspect candidates, correct questions or answers, repair a region, and verify the result in the same interface. SABRE-Prior is our main example of this framework. The Counting and Spatial extensions show how it can support other tasks.

Why human verification and repair matter. Current image generators and editors sometimes fail to produce the requested visual change. For example, a prompt asking for a horse with ten legs may produce the wrong number of visible legs, and a checking VLM may also miscount them. Human verification checks that the image meets the task requirements and that the reference answer follows from the image. Difficulty screening helps focus this work on promising candidates. In Context, it reduced the number of cases requiring human review from 220 to 147 (33.2%) (Appendix D.4). As generation, editing, and automated checking improve, we expect fewer cases to require manual correction. SABRE's modular design allows these components to be replaced and the review process to be adapted as their capabilities improve.

Construction comparison on cable tracing. We ran both workflows on the same task and used AutoBench-V's highest difficulty setting (hard). We applied the same human verification criteria to both sets. Labels had to be readable, cables had to be traceable through crossings, and the correct endpoint had to be clear from the image. Each workflow produced 20 examples that met these criteria.

MetricAutoBench-V (hard)SABRE
Complete candidates generated9742
Examples accepted after human verification2020
Examples accepted after localized repair and reinspection012
Total batch construction timeNearly 4 hoursAbout 30 minutes
GPT-5.4 accuracy15/20 (75%)1/20 (5%)
Gemini accuracy18/20 (90%)1/20 (5%)
Claude Sonnet 4.6 accuracy16/20 (80%)2/20 (10%)

Both workflows used image-generation APIs. In our AutoBench-V reproduction, images that failed inspection were discarded or regenerated, and each replacement had to be checked again. With SABRE, 12 of the 20 accepted examples (60%) had local defects that reviewers corrected using the repair tool. For example, reviewers could correct a local mirroring error while keeping the parts of the cable layout that already met the task requirements. This allowed us to retain those 12 candidates without generating and inspecting full-image replacements.

In this comparison, SABRE obtained 20 valid examples from 42 complete candidates (47.6%), compared with 20 from 97 (20.6%) for AutoBench-V. It also took less time to build the batch. All three VLMs scored lower on the verified SABRE examples, which is the intended outcome for a stress test designed to expose model failures.

This small-scale comparison measures the two workflows as used, including their own difficulty-selection procedures. The batch times cover generation, screening, review, repair or regeneration, and final verification. They provide a direct comparison of elapsed construction time and candidate yield. A controlled estimate of total model-call cost and active human labor per valid challenging example would require separate cost and labor logs, with the foundation models and difficulty-filtering procedure held fixed. We will add the comparison and make its scope clear in the revision.

(2) How to interpret Table 2 and the All4 metric

The reviewer asks how to interpret cross-benchmark score gaps and whether the overall score captures the behavior that SABRE-Prior is designed to test. SABRE is a reusable framework for building targeted stress tests. SABRE-Prior is its main instantiation, designed to test whether VLMs follow visual evidence when it conflicts with learned priors or assumptions suggested by the question.

Table 2 provides cross-benchmark context. The benchmarks in Table 2 all probe related aspects of VLM reliance on priors. We report how the same models perform across these benchmarks using each benchmark's own metric. The score gaps should therefore be read in the context of each benchmark's task design, difficulty filtering, and scoring rules. This presentation follows related work. Table 1 of Vision Language Models Are Biased (Vo et al., ICLR 2026), for example, reports leaderboard scores alongside the question types of related benchmarks. We will make the intended reading of Table 2 clearer in the revision.

All4 measures consistency across a controlled visual change. Consider an office scene with a pen on a desk. A model may correctly say that a pen is present because this matches what it expects to see. We then replace the pen with a crocodile while preserving the surrounding scene. We ask the same two questions about both images:

ImageIs a pen present?Is a crocodile present?
Base imageYesNo
Edited imageNoYes

All4 requires the model to answer all four probes correctly. This tests whether the model can both recognize the newly inserted object and reject the original object after it has been removed. Texture uses the same paired design to test judgments about familiar and unusual materials.

The per-subset and per-probe results show which behaviors drive the overall score. The reviewer's fixed-answer example illustrates why the macro-average should be read alongside these breakdowns. Table 8 reports the individual probe accuracies, which show where models succeed or fail within a paired case. We will point readers to those results more explicitly. We address the Language subset and report the new image-blind results in our response to Question 2.

(3) Case-level retention and preliminary image-quality screening

Thank you for pointing out the inconsistency. We score each question individually and retain candidates at the case level. For Context and Texture, the filtering VLM answers all four questions in a paired case. If it answers all four correctly, we exclude the case from the stress-test candidate pool. If it answers any question incorrectly, we send the entire case to human verification, including the questions it answered correctly. For a single-question case, the rule applies to its one question. We will revise this expression to make this distinction clear.

The preliminary image-quality check uses the structured sample specification. Before difficulty screening, the filtering VLM checks whether the image contains the required visual elements and whether any requested edit appears in the specified region. For example, if a Context specification asks for a bagel to replace a tape roll, the VLM checks that the bagel is visible in the specified region of the edited image and that the tape roll is no longer there. We reject candidates with obvious mismatches and send the remaining candidates to difficulty screening. We will state these criteria explicitly and include this example in the revision.

The two automated checks reduce the number of candidates that reviewers need to inspect. Final inclusion requires human verification that the image meets the task requirements and supports a clear question and a correct reference answer.

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

SABRE addresses the problem of constructing targeted, reliable VLM stress tests at scale as model capabilities evolve and fixed benchmarks become saturated. Collecting and annotating existing images is labor-intensive and offers limited control over the visual conditions needed to probe specific weaknesses. Generative image construction provides this control, but can introduce defects that invalidate questions or reference answers.

The paper proposes a modular, semi-automated data synthesis and curation pipeline. A Test Primer, comprising a natural-language task design, a data schema, and a question format, is converted into structured sample specifications. Existing image generation and editing models produce candidate images and associated question–answer pairs. A filtering VLM screens obvious generation defects and removes readily solved cases, directing human effort toward challenging candidates. Every retained candidate undergoes human verification, with annotation correction and localized image repair when necessary. Thus, reliability comes from screening and curation rather than guaranteed generation fidelity. The central contribution is the integration of these stages into a reusable benchmark-construction workflow.

The main instantiation, SABRE-Prior, tests whether VLMs follow visual evidence when it conflicts with learned world priors. It contains 600 images and 1,000 questions across Context, Texture, Attribute, and Language Elicitation. Six evaluated VLMs achieve macro-average accuracies of 17.8%–31.3%, averaging 22.6%, with distinct weaknesses across subsets. VCD and SoM do not consistently improve performance on the tested Qwen model. Additional Counting and Spatial pilots, each containing 20 samples, demonstrate preliminary applicability beyond world-prior conflicts. These results support the pipeline's ability to construct challenging evaluations; the scores reflect deliberately selected stress cases and subset-specific metrics rather than general model capability.

Critical strengths

Paired counterfactual stress tests reveal failures that ordinary accuracy can conceal. On familiar scenes, a model can answer correctly using learned expectations without grounding its answer in the image. SABRE-Prior deliberately creates conflicts between these expectations and visual evidence, while the paired probes provide more diagnostic information than low aggregate scores alone. For example, Table 8 shows that GPT-5.4 recognizes the original source object in 99% of Context base images, but correctly rejects that object after replacement in only 3% of edited images, despite recognizing the inserted target in 82%. This pattern is consistent with retaining an unsupported expectation even when the replacement is recognized. Such evaluations are valuable for identifying failures to update answers when visual evidence changes.

Critical weaknesses

Reliability remains dependent on human verification and repair, limiting the evidence for scalability. SABRE uses existing image generation and editing models and does not prevent specification violations at generation time. Automatic screening can miss defects, and an incorrect filtering-model response does not distinguish a genuine model failure from an invalid image, ambiguous question, or incorrect reference answer. Human reviewers therefore remain responsible for establishing final sample validity. The paper acknowledges this semi-automated design and reports that screening reduces the Context candidate pool requiring human review from 220 to 147, with 21 of the 100 retained edited images receiving localized repair. However, the reported review and repair times are experience-based estimates rather than systematically timed measurements, and the initial image-output cost excludes several pipeline stages. These statistics provide useful operational evidence but do not establish an end-to-end efficiency advantage over existing counterfactual benchmark-construction workflows under comparable quality and difficulty requirements (Section 3.4–3.5; Appendix D.4).

Questions for authors

Can counterfactual samples produced by SABRE also serve as useful training data, improving visual grounding while preserving ordinary visual performance and commonsense knowledge? Benefits demonstrated by other counterfactual-training methods do not establish the training utility of this particular pipeline. A useful extension would compare equal-budget training on ordinary examples, counterfactual examples, and matched ordinary/counterfactual pairs mixed with ordinary data. Evaluation should include unseen object categories, edits produced independently of the training generator, ordinary real-image tasks, and commonsense questions. Reduced canonical-answer errors that transfer across these conditions, without degrading ordinary accuracy or world knowledge, would provide stronger evidence than improvements confined to similar synthetic images. This is an opportunity to extend the contribution; effective training data are not a prerequisite for a useful evaluation benchmark.

4 comments

(1) Human verification and scalability. Thanks the reviewer asking how SABRE scales when human reviewers still need to verify and sometimes repair generated examples. We agree that a wrong answer from the filtering VLM can reflect a genuine model failure, a defective image, an ambiguous question, or an incorrect reference answer. Human reviewers check the image, question, and answer before an example is accepted.

Building a stress test requires examples that are both valid and challenging. A correctly labeled image can still be easy for a VLM, and images with specific visual changes can be hard to find in existing collections. A manual workflow can therefore require repeated rounds of searching, labeling, and testing before enough suitable examples are found.

SABRE generates candidates from task specifications and screens them for difficulty before human review. This directs reviewers toward candidates that already expose a potential model failure. In Context, screening reduced the number of cases requiring human inspection from 220 to 147, a 33.2% reduction (Appendix D.4).

The review platform brings the images, intended edits, regions to inspect, questions, reference answers, and model predictions into one interface. Reviewers can check the evidence, correct annotations, and record decisions there. When an image has a local defect, they can mark the region, request a repair, and verify the result while keeping the parts of the image that already meet the task requirements. In the final Context subset, 21 of the 100 retained edited images underwent localized repair.

SABRE reduces the number of candidates reviewers need to inspect and gives them a way to correct local defects within the same workflow. We quantify how this workflow performed in a head-to-head construction comparison in (2).

(2) End-to-end construction efficiency and cost. Thanks the reviewer asking how SABRE compares with an existing workflow when both aim to produce valid, challenging examples. We ran a small-scale comparison with AutoBench-V on the same cable-tracing task, using its highest difficulty setting (hard). We applied the same human verification criteria to both sets. Labels had to be readable, each cable had to be traceable through crossings, and the answer had to be clear from the image.

MetricAutoBench-V (hard)SABRE
Complete candidates generated9742
Examples accepted after human verification2020
Examples accepted after localized repair and reinspection012
Total batch construction timeNearly 4 hoursAbout 30 minutes
GPT-5.4 accuracy15/20 (75%)1/20 (5%)
Gemini accuracy18/20 (90%)1/20 (5%)
Claude Sonnet 4.6 accuracy16/20 (80%)2/20 (10%)

SABRE obtained 20 valid examples from 42 candidates (47.6%), compared with 97 candidates (20.6%) for AutoBench-V. Once the examples passed human verification, we used model accuracy to assess their difficulty. All three VLMs scored lower on the SABRE set, so the examples built in less time also exposed more model failures.

The batch times include generation, screening, human review, repair or regeneration, and final verification. Both workflows used image-generation APIs, and individual generation calls accounted for a small share of the time. Much of the work in our AutoBench-V reproduction went into inspecting candidates, regenerating images that failed review, and checking their replacements. With SABRE, 12 of the 20 accepted examples (60%) passed verification after local defects were corrected. For those candidates, reviewers could repair and recheck the affected region without generating a new full image.

Appendix D.4 and Table 7 report 44.44ininitialimage-outputcostsforall220Contextcandidates,orabout0.44 per accepted case. This estimate covers the initial image outputs. The appendix also reports that reviewers typically took under 5 seconds to inspect a candidate and record an initial decision, and under 15 seconds to mark a repair region and enter an instruction. We did not systematically log total active human time per accepted example or the full cost across pipeline stages.

We will add the cable-tracing comparison to the revision and distinguish the observed batch construction times from the estimated image-output costs and typical reviewer operation times.

Question for Authors. Can SABRE-generated counterfactual examples improve visual grounding while preserving ordinary visual performance and commonsense knowledge?

We are glad the reviewer raised this question. SABRE already generates images, questions, and reference answers that can be used for supervised fine-tuning. To build a larger training set, we could skip per-example human verification and tolerate some noise in the generated images or answers. We would use a separate, human-verified test set to measure the effects of training.

Our hypothesis is that counterfactual examples can help VLMs rely more on visual evidence by breaking familiar co-occurrences. For example, pairing an ordinary scene with an edited version that replaces an expected object gives the model a reason to update its answer when the image changes. Training on such pairs may reduce errors caused by relying on familiar object–scene associations when they conflict with the image.

We appreciate the reviewer's suggestion to test this with equal-budget comparisons among ordinary examples, counterfactual examples, and matched ordinary–counterfactual pairs mixed with ordinary data. We would evaluate on unseen object categories and independently generated edits to test whether improvements transfer, and on ordinary real-image tasks and commonsense questions to check whether those abilities are preserved. This is a promising next step for investigating whether SABRE can help address the failures it reveals.

Dear Reviewer ScFN,

Thank you for your careful review and for supporting the acceptance of our paper. We especially appreciated your suggestion to explore whether SABRE-generated examples could improve visual grounding through training.

We know the ICLR discussion period is busy, and we are grateful for the time you have already given our work. If you have a chance to read our response, we would be glad to hear whether it addresses your questions about construction efficiency and whether there is anything else we can clarify. Your feedback has helped us improve the paper.

Area_Chair_7C6HOpenReview ↗

Reviewer feedback request

Dear Reviewer,

Please kindly go through the authors' rebuttal and provide timely feedback. The discussion phase ends in 24 hours (11/18 23:59 AoE). If you have further concerns and need additional clarifications, please specify them and take the limited time the authors have into account. Thank you.

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.