acceptodds
Under review as a conference paper at ICLR 2027

EnterpriseWorld: A Realistic World for Enterprise Long-horizon Agents

Abstract

In enterprise workflows, agents must carry the results, records, and state changes of earlier operations into later ones, often across systems. Existing benchmarks increasingly capture complex, stateful and long-horizon tasks, but rarely make the dependencies between tasks explicit, verified and controllable, limiting both diagnosis of workflow failures and controlled evaluation of changing task relationships. We therefore introduce EnterpriseFlow, a programmatic framework for constructing enterprise workflows around four design principles: realistic enterprise operations, long-horizon coordination, structural control over task dependencies, and verifiable local and cross-task outcomes. Using this framework, we construct EnterpriseWorld, a benchmark of 200 cross-system workflows with 20 operations each in a shared executable environment. Dependencies are explicitly grounded in results and state changes produced by earlier operations, verified by tracing input origins and testing whether upstream errors affect downstream success, and further supported by a human audit with 97.2% agreement. We also construct paired workflow variants that either modify the required dependency topology or add task/tool distractors while preserving it. Our evaluation measures local outcomes, progress along valid dependency chains, and workflow completion, enabling process-level diagnosis of where execution breaks down. Across five frontier models, GPT-6 Astra completes only 46.5% of full workflows, while completion drops by 32-46 points as execution extends across longer dependency chains. Ablation analysis further shows that supplying the correct intermediate results from earlier operations in the dependency chain helps substantially more than exposing the dependency structure, with most dependency failures originating when those results are produced; matched perturbations also break 14-26% of previously successful workflows.

Primary area:
datasets and benchmarks
Keywords:
Benchmark, Long-horizon Reasoning, Enterprise Task

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_7ZuKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EnterpriseWorld contains 200 enterprise workflows with 20 operations each and executable dependency checks. The study separates local correctness from workflow completion and uses oracle information and matched perturbations to examine error propagation.

Critical strengths
  1. Dependencies are tested by corrupting upstream inputs, with additional untouched-state, damage, and skipped-operation controls in Appendix D.3. This makes the workflow construction concrete and verifiable.

  2. Section 4.4 and Table 14 usefully locate many failures in upstream computation rather than information transfer alone. The paper also acknowledges ordinary error accumulation and the extra work introduced by path deepening.

Critical weaknesses
  1. I would like stronger evidence of realism. Verifier sensitivity alone does not establish a plausible business requirement. Human validation helps, but annotators see model verdicts and reasons, and Table 3's workflow-level interval is wide, [71.7%, 98.9%]. Please also compare the construction directly with WorkArena++ [W25], ToolSandbox's state dependencies [TS25], and ResiliBench's stochastic workflow failures [RB26], specifying what the dependency interventions add.

  2. The 14–26% perturbation break rates lack an unchanged-workflow repeat for comparison. With one attempt per cell, some failures may reflect ordinary variability. The paper acknowledges this limitation, but a repeat reference is needed to interpret the magnitude. I also could not locate the bidirectional transition counts referred to in Appendix D.

I lean toward acceptance for the executable dependency construction and process analysis. Stronger realism validation and a repeat reference would improve the conclusions.

Questions for authors
  1. How many workflows were human-reviewed, how were they sampled, and how were questionable compositions handled? A short comparison with [W25, TS25, RB26] should distinguish state dependence, structural perturbations, stochastic failures, and scoring.

  2. Is an unchanged-base rerun available for the perturbation subset? Please compare break rates and provide both success-to-failure and failure-to-success counts.

[W25] Boisvert, L., Thakkar, M., Gasse, M., Caccia, M., Le Sellier De Chezelles, T., Cappart, Q., Chapados, N., Lacoste, A., and Drouin, A. (2025). WorkArena++: Towards Compositional Planning and Reasoning-Based Common Knowledge Work Tasks. arXiv:2407.05291, version 2. https://arxiv.org/abs/2407.05291.

[TS25] Lu, J., Holleis, T., Zhang, Y., Aumayer, B., Nan, F., Bai, H., Ma, S., Ma, S., Bao, M., Yin, G., Wang, Z., and Pang, R. (2025). ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities. Findings of the Association for Computational Linguistics: NAACL 2025, 1160–1183. https://doi.org/10.18653/v1/2025.findings-naacl.65.

[RB26] Ao, R., Min, Z., Zhu, T., Yin, W., and Wang, X. (2026). ResiliBench: Evaluating Agentic Workflow Adaptation in Stochastic Environments. International Conference on Learning Representations. https://openreview.net/forum?id=KTZ56LG7jZ.

6 comments

Response to Reviewer 7ZuK

Thanks for your insightful review. We appreciate your recognition of the value of our work!


Response to Concern of Realism

I would like stronger evidence of realism.

We sincerely thank the reviewer for raising this concern. AI verdicts and rationales were shown to facilitate annotation, but annotators were explicitly informed that these judgments could be incorrect, and the AI model's identity was concealed. The reported audit randomly sampled 30% of retained workflows, and full-dataset annotation is ongoing; we will update the statistics accordingly.

Our realism is grounded in plausible enterprise business processes, where calculations constrain approvals, payments, and reconciliations across shared records. Beyond verifier sensitivity, human annotation explicitly assesses workflow coherence, business justification, missing steps, and quantity mappings.

Regarding related work, WorkArena++ evaluates composed ServiceNow tasks, ToolSandbox evaluates state-dependent tool use through milestone DAGs, and ResiliBench examines workflow resilience under stochastic tool failures. EnterpriseWorld instead emphasizes counterfactually verified dependencies and matched structural interventions that deepen dependency paths or add prerequisites, enabling controlled evaluation of changing coordination requirements rather than stochastic execution failures.


Response to Concern of Perturbation Repeat

The 14–26% perturbation break rates lack an unchanged-workflow repeat for comparison. Is an unchanged-base rerun available for the perturbation subset?

We sincerely thank the reviewer for raising this concern. We conducted two additional repetitions of the unchanged Base and all six perturbations on the same 100 workflows using GPT-5.6 Sol. Each perturbation is paired with its corresponding Base run. We report the success-to-failure (S→F) transitions across all three runs and the original failure-to-success (F→S) counts.

SettingOriginal S→FRepeat 1 S→FRepeat 2 S→FOriginal F→S
Path deepening (k=1)7/439/414/397/57
Path deepening (k=3)11/438/415/3911/57
Path deepening (k=5)10/438/413/397/57
Prerequisite addition6/435/414/399/57
Task distractors (+10)8/438/416/3910/57
Tool distractors (+10)9/4312/414/397/57

Unchanged Base-to-Base comparisons show 7.7–20.9% natural break rates, with both S→F and F→S transitions. Despite this variability, all six perturbations consistently break previously successful workflows across all three runs.


Response to Human Annotation

How many workflows were human-reviewed, how were they sampled, and how were questionable compositions handled?

We sincerely thank the reviewer for raising this concern. The human audit reported in the paper is based on a randomly selected 30% sample of the workflows retained after data filtering. We are also continuing human annotation across the full dataset. As this annotation progresses, we will update the audit results with the expanded coverage, explicit workflow counts, and corresponding annotation statistics.


Response to Literature Comparison

A short comparison with [W25, TS25, RB26] should distinguish state dependence, structural perturbations, stochastic failures, and scoring.

We sincerely thank the reviewer for raising this concern. We clarify the differences in state dependence, workflow interventions, stochastic failures, and scoring below. Here, stochastic faults refer to injected tool failures.

BenchmarkState dependenceWorkflow variationStochastic faultsScoring
WorkArena++ [W25]Application stateTask composition—Task validators
ToolSandbox [TS25]Tool prerequisitesState/tool variants—Milestone DAG
ResiliBench [RB26]Tool prerequisitesFlawed instructions✓Execution criteria
EnterpriseWorldValues, records, statesMatched dependency edits—Node/edge + closure

EnterpriseWorld differs by combining counterfactually verified dependencies with matched structural interventions that deepen dependency paths or add prerequisites. Its node/edge and closure checks further distinguish local correctness from dependency-valid completion, enabling controlled evaluation of changing workflow coordination requirements.

Hi, just a quick reminder that we’ve posted our rebuttal. Looking forward to your feedback, and thanks again for taking the time to review!

Thank you for the response and the reruns. The comparison table answers my related-work question. The response does not say whether it will be added to the paper, and the submission covers only part of this line of benchmarks, so the positioning still seems incomplete for a benchmark paper. I would encourage the authors to include the comparison.

My main concern was whether the perturbation break rates exceed rerun variability. The new counts suggest they do not. In the original run the six perturbations give 51 success-to-failure and 51 failure-to-success transitions in total, so variant success stays between 40 and 46 out of 100 against Base 43/100. The reported Base-to-Base break rates (7.7% to 20.9%) also overlap the perturbation rates. I am therefore not convinced that the 14% to 26% figures in the abstract show fragility rather than run-to-run churn, and Finding 1 would need to be restated.

The audit is described as 30% of workflows, but no count is given, and the workflow-level interval in Table 3 appears to correspond to 15 of 16 judgments. Because the perturbation result is a headline finding and the main point of difference from prior benchmarks, I have lowered my score to 2.

Reviewer_U2svOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper gives EnterpriseWorld, a benchmark of 200 enterprise workflows with 20 tasks each, built from existing benchmarks in one shared SQLite world. Each dependency between tasks is explicit and tested, and the authors also make variants that change the dependency structure or add distractors. On five frontier models, the best model completes 46.5% of workflows, and most failures start when an early task makes a wrong value.

Critical strengths
  • The paper moves the evaluation from final task success to the level of each dependency. Each edge is tested with a wrong input, and each task has its own executable check. Thus, the benchmark can show where a workflow first fails, and most end-state benchmarks cannot do this.
  • The appendix gives a lot of detail on construction, validation gates, and scoring. This makes the benchmark easier to examine and reproduce.
Critical weaknesses
  • The agent has only six generic SQLite tools, and each “system” is only a table in one database. Many tasks are BIRD SQL questions with a business story added. The paper does not show that these workflows are similar to real user requests.
  • The source tasks and databases come from existing benchmarks, and the new part is mainly the tasks that connect them. AutomationBench and Toolathlon already test long workflows across many apps, and the paper does not cite Toolathlon. Thus, it is not clear what this benchmark measures that the others do not.
  • Most findings are generally expected, and the paper’s own independence baseline explains much of the gap between local accuracy and workflow survival. This suggests that most failures come from local errors, not from the dependency structure itself. The paper also has many custom metrics, but they mostly describe failures and give few clear lessons for agent design.
Questions for authors

See the weakness section above

4 comments

Response to Reviewer U2sv

Thanks for your review.


Response to System Settings

The agent has only six generic SQLite tools, and each “system” is only a table in one database.

We sincerely thank the reviewer for raising this concern. The six SQLite tools provide a standardized interface for reproducible execution and dependency verification, rather than reducing workflows to SQL queries. Beyond BIRD source lookups, EnterpriseOps-Gym, WorkBench, and AppWorld contribute CRM, calendar, payment, and communication operations. Agents must create/update records and satisfy cross-operation requirements, verified independently of source-native checks.

Importantly, 44.7% of dependency edges carry business state, and 2,248 of 4,000 operations are state-sensitive. The balance-transfer case illustrates how an incorrect computation changes the review branch and affects subsequent ledger postings, handoffs, and CRM updates, demonstrating executable consequences beyond SQL retrieval.

Business plausibility is further assessed through human annotation of workflow coherence, business justification, missing steps, and quantity mappings, with 93.8% endorsement of sampled workflow-level judgments.


Response to Task Settings

The source tasks and databases come from existing benchmarks, and the new part is mainly the tasks that connect them.

We sincerely thank the reviewer for raising this concern. Unlike benchmarks primarily evaluating long-horizon execution across applications, EnterpriseWorld focuses on explicit, verifiable, and controllable cross-operation dependencies. Reusing validated source tasks is deliberate: we preserve their native executable checks while introducing case-specific value/state dependencies. Each dependency is counterfactually verified by showing that an incorrect upstream input causes downstream verification to fail.

While AutomationBench and Toolathlon also evaluate cross-application workflows, EnterpriseWorld introduces (1) matched dependency interventions that modify required paths and prerequisites, alongside task/tool distractor controls, and (2) dependency-aware diagnostics that distinguish local correctness from dependency-valid completion and locate producer, handoff, and consumer failures. These capabilities enable controlled evaluation of changing coordination requirements and precise failure localization, beyond measuring overall workflow success.


Response to Paper Findings

Most findings are generally expected, and the paper’s own independence baseline explains much of the gap between local accuracy and workflow survival.

We sincerely thank the reviewer for raising this concern. While the independence baseline captures accumulated local errors, it does not reveal their downstream consequences or actionable interventions. Our analysis provides three concrete insights: (1) Verify intermediate results before reuse: the Handoff Oracle improves LHCR-Req by 22.8 -- 29.5%, compared with 6.8 -- 10.5% for the Dependency-Map Oracle, with most dependency failures originating at producers. (2) Recheck prerequisite state: 60–79% of state-sensitive failures with correct prerequisites involve stale-state/binding errors. (3) Adapt procedural guidance: the Procedure Oracle improves GPT-5.6 Sol by 21.5% but changes GPT-6 Astra by −0.5%.

To further examine dependency-related failures, we find that standalone accuracy drops to ~10% when required prerequisites are withheld, making dependent tasks underspecified. We also decompose existing trajectories into primary failures (correct ancestors), downstream failures (incorrect own-checks and ancestors), and ancestor-blocked operations (correct own-checks but invalid dependency closure).

ModelPrimary failure (%)Downstream failure (%)Ancestor-blocked (%)
GPT-6 Astra3.612.67.2
GPT-5.6 Sol3.89.710.7

Percentages are over required operation-runs. Over 84% of dependency-invalid operations have incorrect ancestors, including 7.2–10.7% of required operations that pass local checks but fail dependency closure. Thus, beyond measuring error accumulation, EnterpriseWorld reveals its downstream consequences and identifies concrete priorities for improving agent reliability.

Hi, just a quick reminder that we’ve posted our rebuttal. Looking forward to your feedback, and thanks again for taking the time to review!

Thank you for the response. The additional analysis addresses some of my concerns. However, my main concern remains the benchmark’s practical value, given that it mostly combines tasks from existing benchmarks. I’m keeping my original score.

We sincerely thank the reviewer for the follow-up. Reusing validated tasks from existing benchmarks is a deliberate design choice that enables scalable workflow construction while preserving source-native checks. Our contribution goes beyond task aggregation: EnterpriseWorld introduces counterfactually verified cross-operation dependencies, controllable structural interventions, and dependency-aware diagnostics, which existing source benchmarks do not provide.

Reviewer_7yQDOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

EnterpriseWorld is a benchmark for long-horizon agents built from 200 enterprise workflows with 20 operations each. Its core contribution is the EnterpriseFlow framework that combines tasks from existing benchmarks like BIRD, AppWorld and EnterpriseOps-Gym into one shared environment and connects them with case specific tasks, so later steps actually depend on results/records from earlier ones. Each dependency is verified by giving the wrong upstream input and checking that the downstream task fails, so there are no decorative dependencies. The agent only gets the goal, case files and tools and the rules only show up as notes once the agent starts working on the related tables, so it has to figure out the procedure by itself. A workflow only counts as solved if the final step + every step it depends on pass, which stops agents from getting credit by just guessing the final answer. They also make variants with longer dependency chains, extra prerequisites and distractor tasks/tools.

Critical strengths
  1. Dependency verification: The way dependencies are built and checked is the core strength of this benchmark. Corrupting inputs to test each dependency, recomputing targets from the live state and negative controls like the guess gate make it hard to pass without doing the actual work.
  2. Meaningful ablations and analysis: The oracle ablations, perturbation variants and depth analysis together give a clear picture of where agents fail. The producer/handoff/consumer split shows most failures start upstream, and the case study is a good example of one early misread breaking the whole workflow.
  3. Clear visuals: The figures, plots and visuals are well made and easy to follow and helped me understand the paper better.
Critical weaknesses

1.Composition gap could be separated more: The paper says agents struggle with combining tasks into a workflow, but most failures happen at the step that first computes a value (the producer), and multiplying the step accuracies together (the independence baseline) already explains most of the drop with depth. So agents seem to mostly fail because single task errors add up over many steps, not because combining tasks is harder. It would be helpful to see results on the same tasks run on their own, if accuracy is similar then the workflow itself isn’t really adding difficulty. 2., Perturbation and human annotation results: The perturbation study is only on GPT-5.6 Sol with one run per workflow, and break rates are computed only over the base successful workflows, so some of the break rate could potentially just be run variance. In the human audit the annotators always see the AI verdict before judging it (Appendix D.4), which could bias them towards agreeing more often, and some CIs are wide (B2 is [74.1, 100.0] and the workflow-level questions are [71.7, 98.9]).

Questions for authors

A. Clarifications 1.Human audit: How many workflows were in the human audit? I couldn’t find the number. Oracles: Are the oracles in Table 1 added on top of each other or each run separately on the default? The caption says additive but Appendix E.3 reads like separately.

B. Additional analysis about Agent behavior

  1. Desk notes: In the case study the note with the right definition shows up at call 47 but the agent still adds the wrong filter at call 53. How often do agents get the rule and still not follow it?
  2. Fixing mistakes: Do agents ever go back and fix an earlier wrong write, or do they mostly build on whatever they did first like in the case study? 3 Irreversible steps: About 10% of operations are irreversible. Do agents fail these more often, or are they more careful with them?
3 comments

Response to Reviewer 7yQD

Thanks for your insightful review. We appreciate your recognition of the value of our work!


Response to Composition Gap and Standard Evaluation

Composition gap could be separated more

We sincerely thank the reviewer for raising this concern. EnterpriseWorld focuses on constructing, verifying, and modifying consequential dependencies, testing agents' ability to maintain correct intermediate results, states, and prerequisites across operations. Standalone accuracy drops to ~10% when prerequisites are withheld, making dependent tasks underspecified; providing correct prerequisites instead bypasses the coordination requirements we aim to evaluate.

To distinguish local errors from their downstream consequences, we further decompose existing trajectories. Primary failures have correct ancestors; downstream failures have incorrect own-checks and ancestors; ancestor-blocked operations pass locally but fail dependency closure.

ModelPrimary failure (%)Downstream failure (%)Ancestor-blocked (%)
GPT-6 Astra3.612.67.2
GPT-5.6 Sol3.89.710.7

Thus, while the independence baseline captures error accumulation, EnterpriseWorld reveals its consequences for downstream dependency validity, which isolated task accuracy cannot capture.


Response to Concerns about Perturbation

The perturbation study is only on GPT-5.6 Sol with one run per workflow, and break rates are computed only over the base successful workflows, so some of the break rate could potentially just be run variance.

We sincerely thank the reviewer for raising this concern. To assess run-to-run variability, we conducted two additional repetitions of the unchanged Base and all six perturbations on the same 100 workflows using GPT-5.6 Sol. Each perturbation is paired with its corresponding Base run.

SettingOriginal (a1)Repeat 1 (a2)Repeat 2 (a3)
Base successes43/10041/10039/100
Path deepening (k=1)7/439/414/39
Path deepening (k=3)11/438/415/39
Path deepening (k=5)10/438/413/39
Prerequisite addition6/435/414/39
Task distractors (+10)8/438/416/39
Tool distractors (+10)9/4312/414/39

Despite run-to-run fluctuations, the qualitative finding is consistent: all six perturbations break previously successful workflows across all three runs, with break rates ranging from 7.7% to 29.3%.


Response to Human Annotation pipeline

In the human audit the annotators always see the AI verdict before judging it (Appendix D.4), which could bias them towards agreeing more often, and some CIs are wide (B2 is [74.1, 100.0] and the workflow-level questions are [71.7, 98.9]).

We sincerely thank the reviewer for raising this concern. We presented the AI verdict and its rationale to help annotators understand the proposed dependencies and their business context more efficiently. Before annotation, we explicitly informed annotators that the AI verdicts could be incorrect and should not be treated as ground truth. We also concealed the identity of the AI model from annotators.


Response to Human Annotation Statistics

Human audit: How many workflows were in the human audit? I couldn’t find the number.

We sincerely thank the reviewer for raising this concern. The human audit reported in the paper is based on a randomly selected 30% sample of the workflows retained after data filtering. We are also continuing human annotation across the full dataset. As this annotation progresses, we will update the audit results with the expanded coverage, explicit workflow counts, and corresponding annotation statistics.


Response to Oracle Evaluation

Oracles: Are the oracles in Table 1 added on top of each other or each run separately on the default? The caption says additive but Appendix E.3 reads like separately.

We sincerely thank the reviewer for raising this concern. Each oracle is evaluated separately on top of Default; the oracles are not applied cumulatively. The Procedure Oracle provides the complete procedure and applicable rules upfront. The Dependency-Map Oracle provides upstream dependency relations and their types. The Handoff Oracle provides correct intermediate values or records obtained from the reference execution for the corresponding case.


Response to Error Recovery and Irreversible Operations

Do agents ever go back and fix an earlier wrong write, or do they mostly build on whatever they did first like in the case study?

We sincerely thank the reviewer for raising these questions. Agents can revisit incorrect writes and perform compensating actions without automatic harness repair. We further analyze GPT-5.6 Sol trajectories to quantify error recovery, downstream propagation, and verification behavior across operation types. Here, N counts confirm incorrect writes; repairs are measured at termination, and downstream use is counted before correction. Pre-write verification requires an explicit check of critical values, records, or prerequisites.

Operation typeLocal failure (%)Pre-write verification (%)Repaired (n/N)Used downstream (n/N)
Reversible11.736.621/323157/323
Compensatable11.313.212/3628/36
Irreversible18.910.5N/A66/77

These descriptive results suggest that agents do not consistently exercise greater caution for higher-risk operations.

Hi, just a quick reminder that we’ve posted our rebuttal. Looking forward to your feedback, and thanks again for taking the time to review!

Comments on the paper

1 comment

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

We sincerely thank all reviewers and area chairs for their careful evaluation and constructive feedback. We appreciate the recognition of EnterpriseWorld's contributions and the opportunity to strengthen our work during the rebuttal. Below, we summarize the key strengths recognized across the reviews:

  • Dependency-centered benchmark construction. Reviewer 7yQD identified dependency verification as "the core strength of this benchmark," while 7ZuK highlighted its "concrete and verifiable" construction. U2sv also recognized that dependency-level evaluation enables failure localization beyond conventional end-state benchmarks.

  • Meaningful process-level analysis. Reviewer 7yQD praised the "meaningful ablations and analysis," including oracle interventions and failure attribution. Reviewer 7ZuK similarly recognized the value of identifying upstream computation failures and reaffirmed process-level diagnosis as a main contribution.

  • Rigorous verification and reproducibility. Reviewers 7yQD and 7ZuK acknowledged our counterfactual dependency checks, live reference recomputation, and negative controls. U2sv highlighted the detailed construction, validation, and scoring protocols.

During the rebuttal, we addressed reviewers' concerns through additional experiments and clarifications:

  1. Composition gap and failure analysis. We added a failure decomposition showing that over 84% of dependency-invalid operations have incorrect ancestors, including 7.2–10.7% of required operations that pass local checks but fail dependency closure.

  2. Perturbation reliability. We conducted two additional repetitions of the Base and all six perturbations, reporting bidirectional transitions and run-to-run variability. We clarified that Broken-Req measures conditional fragility rather than a pure causal effect. Reviewer 7ZuK confirmed that this addressed their main concern.

  3. Human validation and realism. We clarified the audit protocol and sampling procedure. Full-dataset human annotation is ongoing and will be included in the final release.

  4. Additional analyses and related work. We added analyses of error recovery, irreversible operations, and downstream propagation, clarified the oracle settings, and provided comparisons with WorkArena++, ToolSandbox, ResiliBench, AutomationBench, and Toolathlon for inclusion in the revision.

Overall, these discussions reinforce EnterpriseWorld's central contribution: making consequential dependencies explicit, verifiable, and controllable to evaluate and diagnose long-horizon workflow reliability beyond individual task success. We sincerely thank all reviewers and area chairs again for their valuable feedback and consideration, and hope the final assessment will consider these contributions alongside the additional evidence and clarifications provided.

Best regards,

All Authors of Submission #19151

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.