acceptodds
Under review as a conference paper at ICLR 2027

VISTA: Looking Right Is Not Working Right — Benchmarking Coding Agents from Design Handoffs to Full-Stack Web and Mobile Apps

Abstract

Coding agents can now build applications from design mockups, but a screen that looks right is not an application that works. Existing benchmarks either score visual reconstruction without a backend or test full-stack functionality from textual requirements, and few tie scores to the individual components a design requires. We introduce VISTA (VIsual Spec-To-App), an end-to-end benchmark in which coding agents turn multi-page design handoffs (Figma renders, structure, and textual requirements) into runnable full-stack Web and Mobile (Android) applications. VISTA provides 8,487 human-annotated interactive components across 18 applications (126 Web pages and 576 Android screens) and an executable evaluator that locates annotated components in the running application and probes their interactions, yielding localization, behavior, and joint scores traceable to individual design requirements. Across 14 deployed coding-agent systems, the best joint score is 0.553 on Web and 0.378 on Mobile, and 13 of 14 systems score lower on behavior than on localization on Web. Component probes detect unresponsive controls in 90.1% of audited Web deliveries. Development traces show where the process departs from the task: 44.5% of trajectories never open the design screenshots, 73.0% omit the prescribed self-audit, and 8.3% report success after a failed check. VISTA makes the gap between rendering a design and delivering a working application measurable at the level of individual components.

Primary area:
datasets and benchmarks
Keywords:
Coding Agents, Multimodal Code Generation, Design-to-Code, Agent Evaluation, Benchmarking

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SnX4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VISTA, a benchmark for coding agents to build Web and Android apps from Figma designs and text requirements. There are human annotations for individual components, and it is evaluate whether generated apps have components in the correct location and have the intended functionalities.

Critical strengths
  • Evaluating whether generated apps match the intended design and work correctly is important and needed as more apps become LLM-generated.

  • The human component annotations are a good resource for evaluation.

  • The benchmark covers both Web and Android.

Critical weaknesses
  • The main concern is whether the benchmark can accurately evaluate functionality. The component evaluation checks whether individual components accept input, change state, or navigate. As mentioned in the strengths section, these are useful component tests. However, what is perhaps equally as important are user journey tests for judging whether an app supports the functionality that it was designed for. These could span many components and many different screens to see that all states and data is correctly passed to and from each other. It is unclear whether the current evaluation can accurately checks whether user journeys are correctly implemented.

  • Similarly, the human annotation and validation (which again are useful) also does not cover these important user journeys and whether the generated apps correctly support these journies.

Questions for authors
  1. Can we safely assume that if components work as they should, multi-component & multi-screen user journeys would also work as expected?

  2. How are backend updates, persistence, and consistency across pages verified?

4 comments

We thank the reviewer for recognizing the importance of evaluating design fidelity and functionality, the value of the component annotations, and the coverage of both Web and Android. Your suggestion to evaluate complete user journeys motivated additional workflow tests, which we describe below.

1. Can we safely assume that working components imply successful multi-component and multi-screen user journeys?

User journeys build on sequences of component interactions, making the component annotations a foundation for workflow-level evaluation. To evaluate how these interactions compose across screens, we added journey tests that execute annotated components in sequence and verify the resulting state and data consistency.

Specifically, we added 18 user-journey definitions across nine task categories, evaluated on four systems using 36 original generated Web applications and 72 workflow executions. These tests check whether components work together to complete the selected tasks.

CategoryUser journeys
EcommerceCheckout; cart removal
ForumPublish and reply; voting and user vote view
Newsletter/blogComment; subscribe
ChatDirect message; group message
Real estateProperty inquiry; agent inquiry
Job boardSave applicant profile; job search/filter/detail
TravelReservation; package-tab consistency
Cloud storageUpload/transfer/stored bytes; file search across views
MusicActual playback/pause; save profile
SystemExecutionsPassFail
grok-4.518126
gpt-6-astra18153
opus-4.818711
deepseek-v4-flash18810
Total724230

The overall passing fraction was 58.3%, with an application-clustered 95% confidence interval of [44.4%, 72.2%]. These tests directly assess workflow completion alongside the component-level measurements. We will include the journey definitions, verification protocol, and results in the updated appendix.

2. How are backend updates, persistence, and consistency across pages verified?

The added journey tests verify outcomes at multiple levels:

  • Cross-page consistency: Sequential interactions check whether state and data are correctly carried across components, screens, and views—for example, publishing a reply and retrieving it, or uploading a file and finding it in another view.
  • Backend correctness: For workflows involving backend writes, independent read-only database queries check the expected business records, fields, and relationships.
  • Persistence after restart: All 32 passing server-side workflow executions passed database checks after restarting the backend and database.
  • Recovery in a fresh session: Of these, 18 executions also passed retrieval checks in fresh browser sessions without reused storage state, with re-authentication where required.
  • Operation-specific outcomes: Upload tests verify the stored file bytes, and playback tests require actual media playback.

These journey results are reported separately from the component-level score SS. The component annotations support both levels of evaluation: individual probes assess each interaction, while the journey tests assess their composition across screens and verify the associated backend state and persistence.

I thank the authors for adding the user journey tests. These are good initial analysis that makes the benchmark more complete. I do strongly believe that these more complete sequences of tests are needed in the age of LLM agents. I encourage the authors to continue the effort to cover both web and mobile apps come completely.

However, I think it would be useful to compare the component and journey results on the same apps. For example, how often did all the components in a journey pass their individual tests while the journey itself failed? These sort of analyses make the benchmark much more useful. Also since the new tests only cover web apps, I am unsure whether the same checks can be done for mobile.

Thank you for this suggestion. We have added a paired component--journey analysis on the same generated applications and extended the journey tests to Android. The tests compose existing annotated interactions into workflows and check their resulting cross-screen and business state. Overall, 42/72 Web journeys and 5/36 Android journeys passed.

Android scope and results

The Android suite covers four systems and all eight task categories, with 32 generated applications, nine journey definitions, and 36 execution instances. Five journeys passed and 31 failed.

CategoryWorkflows
MedicalEdit profile; save and reopen
FitnessEdit profile; save and read across views
FurnitureIncrease cart quantity; navigate to profile and back
Food deliveryIncrease cart quantity; visit checkout and return
RecipesEdit profile; save and reopen
FinanceCreate expense; read transaction list
ChatSend message; leave and reopen conversation
EcommerceIncrease cart quantity; remove item; cross-page readback
SystemInstancesPassFail
GPT-6-astra954
Fable-5909
Grok-4.5909
Opus-4.8909
Total36531

For each journey, we collect the component scores for its bound annotated controls and ask whether the journey succeeds when every recorded component receives full behavior credit (Bi=1B_i=1). On Web, three of the six such journeys failed; on Android, two of six failed. The complete cross-tabulation is reported below.

PlatformJourney instancesAll bound components Bi=1B_i=1: journey passesAll bound components Bi=1B_i=1: journey fails
Web7233
Android3642

For example, in an Android cart-removal workflow, the delete and navigation components each received full credit in isolation. Independent SQL confirmed that deletion reduced the stored cart from two items to one, but navigating to the profile and back displayed two cart rows. In a Web profile-editing workflow, all bound components received full credit, yet the saved name disappeared when settings were reopened. These cases show why checking individual responses does not establish cross-screen state consistency, and identify failures that the added journey evaluation captures.

We will add the paired tables and representative failure traces to the appendix and include the Android workflow protocol and evidence with the benchmark materials. We will continue expanding the journey suite to cover additional workflows on both Web and Android.

We thank the reviewer for these constructive suggestions, which have helped make the paper clearer and the evaluation more rigorous.

We have further expanded the journey evaluation with 16 additional workflow definitions and 64 additional executions across Web and Android. The updated app.J reports all 172 journey executions, including the paired component–journey analysis and representative failure cases. We appreciate the reviewer’s thoughtful comments and constructive discussion, which have helped strengthen the paper.

Reviewer_k2qTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

VISTA evaluates coding agents that generate web and Android applications from design handoffs, rather than from conventional implementation specifications alone. The benchmark includes ten web apps and eight Android apps, with 126 web pages, 576 Android screens, and 8,487 annotated interactive components. Its principal evaluation combines component localization and behavioral probes, and it examines 14 agent/harness configurations with three attempts per task. The reported best joint scores are 0.553 on web and 0.378 on mobile; workflow analyses also identify failures in screenshot inspection and self-checking.

Critical strengths
  1. A realistic, challenging coding-agent task. Generating interactive deployable applications from visual handoffs is more demanding and practical than image-only reconstruction. The inclusion of web and Android tasks exposes failures that would be invisible in a screenshot similarity benchmark.

  2. More interpretable failure signals than a single visual score. By reporting component localization and behavior separately, and inspecting agent tool traces, the benchmark gives evidence for the important observation that visually plausible implementations can still be functionally incomplete.

Critical weaknesses
  1. The core metric does not convincingly measure full-stack correctness. Most described behavioral probes test interface-level responses, navigation, or visible state changes. It remains unclear how often they independently verify server-side business logic, database persistence, recovery across sessions, or authentic backend effects. A system using mocked or nonpersistent backend behavior could therefore perform relatively well without meeting the claimed end-to-end functionality requirement. This is a construct-validity issue, not merely a missing additional experiment.

  2. The Android scoring denominator and evaluator calibration may bias comparisons. Unlike web scoring, Android excludes components whose behavior cannot be determined from the denominator. Without per-agent missingness rates and a sensitivity analysis treating such cases as failures, the relative rankings are difficult to interpret. Moreover, the reported human behavioral validation covers only 24 probe anchors, which is limited for the platform and interaction diversity represented by 8,487 component annotations.

Questions for authors

Metric validity: Which benchmark probes independently verify backend/database effects (such as persistence across requests or sessions), rather than just a UI reaction? Can the authors report a small but representative end-to-end workflow evaluation against an independent backend oracle?

Sensitivity analysis: How many Android behavioral probes are excluded for each system, and what happens to the main ranking if all undetermined probes receive zero credit?

Calibration and uncertainty: Can the authors provide human agreement by platform and interaction type, as well as confidence intervals clustered by application rather than by component?

3 comments

We thank the reviewer for the thoughtful feedback on metric validity, Android scoring, and evaluator calibration. These comments motivated additional workflow tests with independent database checks, expanded human validation across both platforms, and application-clustered confidence intervals. We address each question below.

1. Metric validity: Which probes independently verify backend/database effects? Can the authors provide representative end-to-end workflows evaluated against an independent backend oracle?

The core score SS evaluates component localization and observable interaction outcomes; it does not directly score database correctness or persistence.

To independently verify backend effects, we added 18 user-journey definitions across four systems and nine task categories, covering 36 generated Web applications and 72 executions. The tests use the original generated applications and organize existing component annotations into sequential workflows that check state and data consistency across components and screens.

For workflows involving backend writes, independent read-only database queries check expected business records, fields, and relationships. All 32 passing server-side workflow executions passed post-restart database checks; 18 also passed retrieval checks in fresh browser sessions without reused storage state, with re-authentication where required. These workflow results are reported separately from SS. File-upload checks verify stored bytes, and music-playback checks require actual media playback.

CategoryWorkflows
EcommerceCheckout; cart removal
ForumPublish and reply; voting and user vote view
Newsletter/blogComment; subscribe
ChatDirect message; group message
Real estateProperty inquiry; agent inquiry
Job boardSave applicant profile; job search/filter/detail
TravelReservation; package-tab consistency
Cloud storageUpload/transfer/stored bytes; file search across views
MusicActual playback/pause; save profile
SystemExecutionsPassFail
grok-4.518126
gpt-6-astra18153
opus-4.818711
deepseek-v4-flash18810
Total724230

The overall passing fraction was 58.3%, with an application-clustered 95% confidence interval of [44.4%, 72.2%]. These tests directly check completed workflows and backend effects beyond visible UI feedback.

2. Sensitivity analysis: How many Android behavioral probes are excluded per system, and what happens if all undetermined probes receive zero credit?

In the Android main experiment, zero behavioral probes were excluded from the scoring denominator for every system. Unmatched components and failed probes receive zero credit and remain in the denominator. Assigning zero credit to undetermined probes therefore leaves all main-experiment scores and rankings unchanged.

The exclusion rule described in the manuscript applies to intermediate evaluations in a separate staged-generation setting, where navigation targets may be scheduled for later stages. We will clarify this distinction and include the staged-generation score curve in the appendix of the updated PDF.

3. Calibration and uncertainty: Can the authors provide human agreement by platform and interaction type, and confidence intervals clustered by application?

We expanded behavioral validation to 128 samples, with 64 per platform. Within each platform, allocation across interaction types was based on the corresponding component proportions. Automated verdicts agree with human judgments on 96.9% of Web samples and 89.1% of Android samples.

PlatformInteraction typeAgreements / SamplesAgreement
WebNavigation40/40100.0%
WebInput7/887.5%
WebClick9/1090.0%
WebToggle6/6100.0%
WebSubtotal62/6496.9%
AndroidNavigation44/5088.0%
AndroidInput4/4100.0%
AndroidToggle6/6100.0%
AndroidClick3/475.0%
AndroidSubtotal57/6489.1%
OverallAll types119/12893.0%

We also computed 95% confidence intervals using 10,000 application-clustered bootstrap resamples. Each resample retains all repeated runs of each sampled application and recomputes the aggregate score. These intervals will be included in Table 2.

Web joint scores

SystemSSApplication-clustered 95% CI
grok-4.60.553[0.472, 0.638]
fable-50.533[0.457, 0.612]
gpt-5.6-sol0.507[0.459, 0.553]
grok-4.50.491[0.400, 0.586]
opus-4.80.487[0.387, 0.585]
deepseek-v4-flash0.484[0.422, 0.552]
gpt-6-astra0.460[0.359, 0.561]
composer-2.50.448[0.354, 0.552]
gpt-5.50.410[0.308, 0.471]
muse-spark-1.2-contributor0.409[0.344, 0.488]
gpt-5.6-luna0.387[0.299, 0.465]
gpt-5.6-terra0.384[0.311, 0.453]
sonnet-4.60.365[0.278, 0.458]
mai-code-1.1-flash0.293[0.232, 0.359]

Android joint scores

SystemSSApplication-clustered 95% CI
gpt-6-astra0.378[0.156, 0.611]
fable-50.270[0.108, 0.438]
grok-4.60.344[0.196, 0.497]
opus-4.80.282[0.138, 0.420]
grok-4.50.364[0.207, 0.541]
gpt-5.6-sol0.228[0.077, 0.404]
deepseek-v4-flash0.219[0.095, 0.354]
composer-2.50.192[0.099, 0.302]
gpt-5.50.125[0.063, 0.188]
sonnet-4.60.115[0.056, 0.199]
muse-spark-1.2-contributor0.110[0.065, 0.155]
gpt-5.6-terra0.184[0.085, 0.289]
gpt-5.6-luna0.247[0.111, 0.389]
mai-code-1.1-flash0.055[0.027, 0.085]

The rebuttal addressed my concerns. I maintain my positive score.

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

This paper show whether coding agents can turn a design handoff into an app that both looks right and works. The authors provide 10 Web apps (126 pages) and 8 Android apps (576 screens), with 8,487 component-level annotations, also with scoring details.

Critical strengths
  1. The idea is practicsal. It separates looking right from working right, on both Web and Android. Screenshot-based design-to-code evaluation misses exactly what this benchmark measures.

  2. Failures start at deployment, for example an empty init.sql or an Activity that cannot launch. Besides, the trajectory analysis explains why agents fail: Codex supports image input, yet it never opens a screenshot in 71.1% of trajectories. In one trace, a set -e check exits because wget is missing, and the agent still reports that its checks passed.

Critical weaknesses
  1. For a benchmark paper, details means many things. The data-testid contract covers roughly 600 elements, against 3,253 Web annotations. It is unclear how the remaining components are matched. It is unclear what happens to components that cannot be found.

  2. Some headline claims go beyond the evidence. Table 2 has no confidence intervals, while the condition-study intervals are about ±0.1. Could the authors check this? Mobile has no human validation, and the joint score correlates with human judgment is low.

Questions for authors

Several questions:

a. Who annotated the components, and what was the inter-annotator agreement?

b. What are the exact definitions of L and B? How are unmatched components handled, and how many are excluded per system?

c. At the component level, what is behaviour accuracy conditional on correct layout? This would test the B < L claim directly.

2 comments

We thank the reviewer for the careful assessment and constructive questions about annotation agreement, component matching, score definitions, and statistical uncertainty. We respond to Questions a–c individually below, followed by the concerns about confidence intervals and human validation.

a. Who annotated the components, and what was the inter-annotator agreement?

Three annotators produced the original component annotations, each responsible for a separate subset. Three additional annotators independently re-annotated the same 64 sampled components without access to the original annotations. They unanimously agreed on 95.3% of components; agreement with the original annotations was 98.4% across 192 comparisons.

PlatformInteraction typeComponentsMatches / ComparisonsAgreement with original annotations
WebNavigation824/24100.0%
WebInput720/2195.2%
WebToggle720/2195.2%
WebClick824/24100.0%
WebScroll13/3100.0%
WebOther13/3100.0%
WebSubtotal3294/9697.9%
AndroidNavigation721/21100.0%
AndroidInput618/18100.0%
AndroidToggle618/18100.0%
AndroidClick721/21100.0%
AndroidScroll617/1894.4%
AndroidSubtotal3295/9699.0%
OverallAll types64189/19298.4%

b. What are the exact definitions of LL and BB? How are unmatched components handled, and how many are excluded per system?

The 3,253 Web interaction annotations provide bounding boxes, interaction types, and expected behaviors. Testids provide sparse alignment anchors; evaluation also covers main-scored components without testids. High-confidence correspondences establish a page-level coordinate transformation. Other annotated components are matched to DOM candidates using aligned bounding boxes, component types, and textual cues.

For each main-scored component ii, LiL_i is determined by its matching tier and BiB_i by the outcome of its interaction-specific probe. For the main-scored set CC, Web scores are:

\bar B=\frac{1}{|C|}\sum_{i\in C}B_i,\qquad S=\frac{1}{|C|}\sum_{i\in C}L_iB_i.$$ Unmatched components receive $L_i=B_i=0$ and remain in the denominator; no components are excluded because they are unmatched. In the Android main experiment, the number of excluded behavioral probes is also zero for every system. **Localization rules** | Matching condition, in priority order | Web | Android | $L$ | |---|---|---|---:| | Anchor correspondence | High-confidence semantic/testid anchor | Matching resource-id/content-description and IoU ≥ 0.3 | 1.0 | | Geometric match | IoU ≥ 0.3 | Compatible type and IoU ≥ 0.3 | 1.0 | | Geometric match | IoU ≥ 0.1 | Compatible type and IoU ≥ 0.1 | 0.6 | | Nearby match | Center distance ≤ 150 px | Compatible type and center distance < 55 dp | 0.3 | | Relaxed match | Center distance ≤ 600 px | Matching identifier without the preceding position conditions | 0.15 | | Final fallback | Reasoning-candidate text similarity ≥ 0.3 | Compatible type and center distance < 200 dp | 0.1 | | Unmatched | No candidate | No candidate | 0 | Web assigns full localization credit to high-confidence anchor correspondences. Android uses one-to-one assignment within each screen. **Behavior rules** | Interaction | Web | Android | |---|---|---| | Input | Full value acceptance: 1; partial acceptance: 0.5; failure: 0 | Correct target-field/password-masked feedback: 1; text appearing elsewhere: 0.5; no feedback: 0 | | Toggle | Change in checked/pressed/expanded, control class, or theme state: 1; otherwise: 0 | Checked state flips: 1; for non-checkable controls, screen-fingerprint Jaccard < 0.97: 1; otherwise: 0 | | Navigation | Visible-content change, with no-op and unintended home-reversion checks | Target is nearest page fingerprint: 1; target in top three or closer than source: 0.5; otherwise: 0 | | Pop-out | Increased dialog/expanded-element count: 1; otherwise: 0 | Dialog-like element: 1; at least three new fingerprint elements: 0.5; otherwise: 0 | | Generic click | URL change: 1; HTML-length change > 50: 0.7; successful click without these changes: 0.5; failed click: 0 | Screen-fingerprint Jaccard < 0.97: 1; otherwise: 0 | | External action | External href: 1, subject to the implemented link check | Foreground package changes: 1; otherwise: 0 | Both platforms evaluate annotated interaction outcomes using platform-specific runtime evidence: DOM and visible-content changes on Web, and UI-hierarchy states and screen fingerprints on Android. The exact thresholds and fallback rules will be documented in the updated appendix. Results are reported separately for each platform.

c. At the component level, what is behaviour accuracy conditional on correct layout? This would test the B<LB<L claim directly.

For components receiving maximum localization credit (L=1L=1), we report both the mean behavior score and the fraction receiving full behavior credit (B=1B=1), with application-clustered confidence intervals.

SystemComponents with L=1L=1Mean BB given L=1L=1P(B=1∣L=1)P(B=1\mid L=1)Application-clustered 95% CI for PP
grok-4.638010.8300.812[0.682, 0.880]
fable-540820.8130.790[0.643, 0.850]
gpt-5.6-sol13420.7730.756[0.675, 0.798]
grok-4.536520.8050.788[0.612, 0.868]
opus-4.836830.7980.782[0.624, 0.859]
deepseek-v4-flash36930.7650.746[0.609, 0.812]
gpt-6-astra23450.7180.696[0.489, 0.831]
composer-2.521620.8370.817[0.678, 0.883]
gpt-5.580980.5570.533[0.442, 0.699]
muse-spark-1.2-contributor23160.8160.787[0.699, 0.859]
gpt-5.6-luna27340.5550.542[0.357, 0.797]
gpt-5.6-terra37260.7410.720[0.516, 0.814]
sonnet-4.621420.6000.546[0.351, 0.663]
mai-code-1.1-flash23940.8240.802[0.650, 0.885]

These results show that receiving maximum localization credit does not guarantee successful behavior.

Additional concern: Confidence intervals for Table 2

We computed 95% confidence intervals using 10,000 application-clustered bootstrap resamples. Each resample retains all repeated runs of each sampled application and recomputes the aggregate score. We will include these intervals in Table 2.

SystemWeb SSApplication-clustered 95% CI
grok-4.60.553[0.472, 0.638]
fable-50.533[0.457, 0.612]
gpt-5.6-sol0.507[0.459, 0.553]
grok-4.50.491[0.400, 0.586]
opus-4.80.487[0.387, 0.585]
deepseek-v4-flash0.484[0.422, 0.552]
gpt-6-astra0.460[0.359, 0.561]
composer-2.50.448[0.354, 0.552]
gpt-5.50.410[0.308, 0.471]
muse-spark-1.2-contributor0.409[0.344, 0.488]
gpt-5.6-luna0.387[0.299, 0.465]
gpt-5.6-terra0.384[0.311, 0.453]
sonnet-4.60.365[0.278, 0.458]
mai-code-1.1-flash0.293[0.232, 0.359]
SystemAndroid SSApplication-clustered 95% CI
gpt-6-astra0.378[0.156, 0.611]
fable-50.270[0.108, 0.438]
grok-4.60.344[0.196, 0.497]
opus-4.80.282[0.138, 0.420]
grok-4.50.364[0.207, 0.541]
gpt-5.6-sol0.228[0.077, 0.404]
deepseek-v4-flash0.219[0.095, 0.354]
composer-2.50.192[0.099, 0.302]
gpt-5.50.125[0.063, 0.188]
sonnet-4.60.115[0.056, 0.199]
muse-spark-1.2-contributor0.110[0.065, 0.155]
gpt-5.6-terra0.184[0.085, 0.289]
gpt-5.6-luna0.247[0.111, 0.389]
mai-code-1.1-flash0.055[0.027, 0.085]

Additional concern: Mobile human validation and joint-score correlation

The reported ρ=0.43\rho=0.43 compares SS with human visual and structural fidelity ratings. The corresponding visual validation is between LL and these ratings (ρ=0.66\rho=0.66). Behavior is assessed separately.

Our additional behavioral validation covers 64 samples per platform, with interaction-type allocation based on the corresponding component proportions. Automated verdicts agree with human judgments on 96.9% of Web samples (62/64) and 89.1% of Android samples (57/64). We will clarify these validation targets and add the Android results to the appendix.

Comments on the paper

1 comment
Area_Chair_kaJBOpenReview ↗

Reminder: Author–reviewer discussion ends today

Dear Reviewers,

As the author–reviewer discussion period ends today (November 18, AoE), please read the authors’ responses, any relevant manuscript updates, and the other reviewers’ comments, and engage with the authors before the deadline.

In particular, please clarify which of your concerns have been addressed and which remain unresolved, and raise any remaining questions promptly so the authors have an opportunity to respond. Even if your assessment is unchanged, a brief follow-up explaining your reasoning would be helpful.

Thank you for your time and thoughtful reviewing!

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.