acceptodds
Under review as a conference paper at ICLR 2027

FATE: First Agent Teams Exam

Abstract

Real-world professional work requires collaboration among coworkers with distinct responsibilities, information, and permissions, yet existing agent benchmarks provide limited coverage of this setting. We introduce First Agent Teams Exam (FATE), a benchmark for evaluating agent teams on realistic professional tasks. Constructed by 32 domain professionals, FATE comprises 112 tasks across 14 domains and 129 occupations, covering both unimodal and multimodal tasks. Agents use tools and communicate under role-specific resource and permission boundaries to produce intermediate artifacts and final deliverables. We evaluate eight LLMs and compare four communication protocols. Current agent teams achieve a best final evaluation score of only 42.0, with limited collaboration gains at substantial computational cost. These gains vary with communication protocols and execution parallelism, while coordination and artifact delivery remain major bottlenecks to reliable teamwork. We hope FATE will provide a foundation for systematically evaluating agent teams and advancing reliable agent collaboration.

Primary area:
datasets and benchmarks
Keywords:
Benchmark, Multi-Agent, Agent Teams, Working Agent

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_4PxROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents FATE benchmark that tests whether teams of LLM agents can complete realistic professional work together. It contains 112 tasks covering 14 domains and 129 occupations, constructed by 32 domain professionals. In each task, every agent takes an occupational role with its own responsibilities, input resources, and permissions, so the team has to communicate and hand off intermediate artifacts to produce the final deliverables. Scoring happens at two levels: individual members' artifacts (ME) and the team's final deliverables (FE), using both rule-based checks and LLM-judged criteria.

The authors evaluate eight LLMs on FATE. They also compare four communication protocols and conduct a matched-budget comparison between single-agent and multi-agent systems across three domains. The best system reaches an FE 42.0. Multi-agent systems use substantially more tokens and take longer to run, but the performance gains are relatively small. However, when compared under the same computational budget, they perform better overall than single-agent systems, except in the Research domain. Point-to-point communication performs best across the three domains tested. Models generally perform worse on multimodal tasks, and the two most common failure types are communication coordination (34%) and output-interface misalignment (32%).

Critical strengths

I like that FATE focuses on professional tasks where agents actually need to work together, rather than just assigning different roles to solve a problem. Agents have access to different files and permissions, and their work has to be passed between teammates to produce the final deliverables. The tasks also use real materials in their original formats. Having domain experts construct and review the tasks, and test the evaluators with valid and invalid outputs, makes the benchmark more convincing.

I also find the separate evaluation of individual members (ME) and final team outputs (FE) useful. For example, in Building, the average ME is 51.0 while FE is only 16.0. Although these scores use different criteria, they suggest that producing reasonable intermediate artifacts does not necessarily lead to a successful final result. Looking at both gives a better picture of how the teams perform than evaluating only the final output.

Critical weaknesses

The experimental results seem inconsistent. Table 2 and Figure 10 report quite different FE scores for GPT-5.6 Terra on the same three domains, even though both appear to use high reasoning (16.1/56.0/35.0 vs. 6.9/72.8/11.7). This makes it hard to judge how stable the results are

I have some concerns about how the evaluation scores are computed. GPT-5.6 Terra serves as both an evaluated model and the LLM judge. While the 92% human agreement is nice result, it is unclear whether this holds across different evaluated models.

Questions for authors

Given that GPT-5.6 Terra is also used as the LLM judge, I wonder if the authors have checked for potential evaluation bias?

Could the authors explain how FE is computed for tasks with different scoring rules? For example, the Tourism task uses a score cap, while the Finance task uses a continuous FinalScore. I am a little confused

4 comments
AuthorsOpenReview ↗

Response 1

Dear Reviewer 4PxR,

Thank you for your careful and insightful review. We appreciate your positive assessment and your recognition of FATE's focus on realistic professional collaboration and its evaluation design. Below, we provide clarifications and additional evidence concerning experimental consistency and evaluation reliability. We sincerely hope these responses address your concerns!

W1:The experimental results seem inconsistent. Table 2 and Figure 10 report quite different FE scores for GPT-5.6 Terra on the same three domains, even though both appear to use high reasoning (16.1/56.0/35.0 vs. 6.9/72.8/11.7). This makes it hard to judge how stable the results are

Thank you for your question! Figure 10 in paper uses a subset of tasks from the three domains. Table 2 in paper reports results for the full task set in each domain. Because they use different sets of tasks, the FE scores can differ even when both use high reasoning effort. Table 1 of this response compares the subset and full task set FE scores for GPT-5.6 Terra across selected domains. The full task set results show consistent improvements with increasing reasoning effort across these domains. We plan to update Figure 10 using the full-task-set results reported in this table. We observe this same upward trend on both the subset and the full task set, which further shows that higher reasoning effort leads to better collaboration.

Table 1: GPT-5.6 Terra FE Scores on Task Subsets and Full Task Sets Across Selected Domains.

DomainSubset FEFull set FE
lowmediumhighxhighlowmediumhighxhigh
Building0.014.26.99.06.1212.1216.2116.79
Investigation58.462.272.870.046.1652.6456.0658.64
Research6.79.411.720.014.1729.3334.5041.33
AuthorsOpenReview ↗

Response 2

W2 & Q1: I have some concerns about how the evaluation scores are computed. GPT-5.6 Terra serves as both an evaluated model and the LLM judge. While the 92% human agreement is nice result, it is unclear whether this holds across different evaluated models. Given that GPT-5.6 Terra is also used as the LLM judge, I wonder if the authors have checked for potential evaluation bias?

We appreciate your insightful comment regarding potential model bias and cross-model consistency, which are important considerations when using an LLM as a judge.

In FATE, the evaluation criteria are based on objective, verifiable facts, which naturally minimizes subjective bias. To further address your concerns, we conducted an additional validation study on 300 randomly sampled criterion instances. We introduced Gemini 3.8 Flash and Grok 4.6 as independent judges alongside GPT-5.6 Terra to evaluate the exact same instances:

(1) The criterion instances are not sensitive to the choice of judge. As shown in Table 2, across all 300 instances, GPT-5.6 Terra achieves 92.00% agreement with human judgments. Meanwhile, the independent judges Gemini 3.8 Flash and Grok 4.6 achieve 93.67% and 95.00% agreement, respectively. Furthermore, all four evaluators (GPT-5.6 Terra, Gemini 3.8 Flash, Grok 4.6, and Human) reach unanimous consensus on 89.67% of the instances, demonstrating that the evaluation outcomes are highly consistent and do not depend on a specific judge model.

(2) There is no evidence of self-preference bias. As shown in Table 3, in the 38 instances where GPT-5.6 Terra evaluates its own outputs, it agrees with human judgments on 92.11% of cases. In the rare 3 cases of disagreement, GPT-5.6 Terra leaned stricter toward itself (2 strict,1 lenient), showing no tendency to inflate its own scores.

(3) Evaluation consistency remains stable across evaluated models. After excluding its own outputs, GPT-5.6 Terra maintains a 91.98% agreement rate with human judgments on the 262 instances generated by other evaluated models. This aligns closely with its overall performance, confirming that the evaluation reliability of GPT-5.6 Terra holds consistently across different target models.

Table 2: Inter-Judge Consistency and Human Agreement (N = 300)

EvaluatorAgreement with HumanCohen's Kappa
GPT-5.6 Terra92.00%0.839
Gemini 3.8 Flash93.67%0.873
Grok 4.695.00%0.900

Table 3: Evaluation Comparison on GPT-5.6 Terra Outputs (N = 38)

EvaluatorAgreement with Human#More Lenient#Stricter
GPT-5.6 Terra92.11%12
Gemini 3.8 Flash94.74%20
Grok 4.697.37%10

Q2:Could the authors explain how FE is computed for tasks with different scoring rules? For example, the Tourism task uses a score cap, while the Finance task uses a continuous FinalScore. I am a little confused

Thank you for your question. Although different tasks use different scoring rules, their scores are converted to the same range from 0 to 1 before the final task score is calculated.

(1) For rule-based evaluation, each item evaluates to either a binary result or a normalized continuous score. For example, in the Finance task, the raw FinalScore is calculated on a [0, 100] scale based on domain criteria and then divided by 100 to map directly to [0, 1].

(2) For LLM-based evaluation, the model delivers a 0 or 1 judgment for each criterion. The final score for that evaluation item is the average of these binary decisions.

(3) The scores are combined using the weights and rules defined by each task. Any score cap is applied after the initial overall score has been calculated. For example, in the Acadia Tourism task, if the safety score(a task-specific metric) is below 0.6, the final score cannot exceed 0.49. Other tasks may use different conditions.

In short, the process is: calculate each evaluation score → convert it to 0–1 → combine the scores → apply optional task-specific score aggregation → obtain the final task score.

Thank you for the clarifications and additional analysis. I appreciate the authors' responses to my questions. However, I will maintain my original score of 3

AuthorsOpenReview ↗

Response to Reviewer 4PxR

We sincerely thank the reviewer for the positive score! We will incorporate the additional analyses into the revised paper. If you have any further questions, please feel free to let us know.

Reviewer_1vRkOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors introduce a new benchmark called FATE: First Agent's Team Exam. Unlike prior benchmarks evaluating agent's collaboration ability, this new benchmark measures LLM agent ability to collaborate and communicate to solve professional tasks. The benchmark is created with the help of 32 experts across various domains, with each task being rigorously checked for quality control. The final 112 tasks are used to evaluate the current state-of-the-art agent systems such as Claude Code, Codex, and more across varying communication protocols. An error analysis reveals that communication remains a big universal challenge for these agentic systems on professional complex tasks.

Critical strengths

The benchmark is constructed using a rigorous quality control process with the help of domain experts. This makes me trust the results and conclusions more. The authors also don't just evaluate a single communication protocol, instead opting to benchmark 4 different communication stategies to identify the best one. In terms of model selection, the authors choose relevant state-of-the-art agentic harnesses to benchmark, which is relatable and informative for a wide audience using these tools.

Critical weaknesses

Some details are missing from some sections. For example, the paper mentions that all private and sensitive information are removed on lines 201/202, but does not elaborate on how it was removed or what constitutes private information.

For the experiment results, there seems to be no confidence interval associated with each model's performance. Given the small number of tasks in the benchmark, there might be a lot of variance in the final evaluated outcome, and some discussion or reporting of this would strengthen the paper.

Questions for authors

In the experiment comparing single versus multi-agent performance, what is the final value chosen for "n"? The paper states that a single-agent performs n rollouts of the task and the best outcome is selected, but does not plot or say the final chosen value for this hyperparameter.

1 comment

Dear Reviewer 1vRk,

Thank you for your careful and insightful review. We are especially grateful for your Clear Accept assessment and your recognition of the rigor and relevance of our work. Below, we clarify the details and provide additional experimental results in response to your questions. We sincerely appreciate your support and constructive feedback!

W1:Some details are missing from some sections. For example, the paper mentions that all private and sensitive information are removed on lines 201/202, but does not elaborate on how it was removed or what constitutes private information.

Thank you for your suggestion. The details of how we handle private and sensitive information are provided below. Before adding source materials to FATE, we check the data for personal identity information, case information, and sensitive project information. This covers direct details such as names, addresses, and contact information, as well as any personal information that appears incidentally in media files such as images, videos, or maps.
As shown in Table 1, we use the Remove, Generalize, and Replace methods to handle private and sensitive information. We remove any information that is not needed for the task. For fields that remain necessary, we replace the original values or generalize them, keeping only the level of detail and data relationships required to complete the task. For example, if a task requires matching records for the same person across multiple files, we replace their real name with a consistent substitute name to preserve the connection between records. If the task only requires knowing the general area where a person lives, we keep the regional information and remove the specific street address.
Table1: Methods for Handling Private and Sensitive Information in FATE.

MethodMeaningObjective
RemoveDelete personal or sensitive information not needed for the taskPrevent unnecessary privacy exposure
GeneralizeConvert specific details into broader categoriesReduce identifiability while retaining context needed for the task
ReplaceSubstitute sensitive values with alternative valuesPreserve data structures and factual relationships required by the task

W2:For the experiment results, there seems to be no confidence interval associated with each model's performance. Given the small number of tasks in the benchmark, there might be a lot of variance in the final evaluated outcome, and some discussion or reporting of this would strengthen the paper.

We thank the reviewer for the suggestion. Table 2 reports the domain-level FE scores and corresponding confidence intervals for selected models. We run the experiment three times, and the results are consistent across runs. In light of the cost of running experiments with interactions that span many turns over a long horizon, we report the main results from a single run in the paper. It does not affect any of the paper's conclusions.
Table2: FE Scores and Confidence Intervals for Selected Models.

DomainGrok 4.6Gemini 3.8 FlashGPT-5.6 Terra
Building18.99 [16.72, 21.26]1.73 [0.00, 3.46]16.64 [15.12, 18.16]
CFD52.38 [46.98, 57.77]37.20 [33.14, 41.26]27.90 [22.02, 33.77]
Education14.55 [10.54, 18.57]16.10 [11.70, 20.49]13.38 [10.53, 16.24]
Investigation62.86 [57.31, 67.07]55.11 [53.52, 56.70]53.31 [48.70, 58.32]
Legal27.49 [26.59, 28.39]10.75 [7.85, 13.65]24.51 [21.63, 27.38]
Research19.63 [17.96, 21.30]3.03 [0.00, 6.06]34.68 [32.24, 37.12]
Health46.32 [42.40, 50.25]48.80 [37.75, 59.85]68.66 [65.36, 71.96]

Q1:In the experiment comparing single versus multi-agent performance, what is the final value chosen for "n"? The paper states that a single-agent performs n rollouts of the task and the best outcome is selected, but does not plot or say the final chosen value for this hyperparameter.

We apologize for the misunderstanding. nn is not fixed across all tasks; it is determined separately for each task. For task ii, we calculate the ratio of token usage in the multi-agent setting to that in the single-agent setting and round it up to obtain the number of repeated runs:
$ n_i=\left\lceil \frac{T_{\mathrm{multi},i}}{T_{\mathrm{single},i}} \right\rceil, $
where Tmulti,iT_{\mathrm{multi},i} and Tsingle,iT_{\mathrm{single},i} denote the token usage for task ii in the multi-agent and single-agent settings, respectively. Therefore, nin_i may vary across tasks.

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

The paper introduce FATE, a benchmark for evaluating agent teams on realistic professional tasks. FATE selects real-world scenarios that require cross-occupation collaboration using the O*NET taxonomy, and then collects multimodal resources to construct the benchmark. For each task, FATE annotates occupational roles, resource access, permissions, and workflow dependencies. The paper evaluates eight model-harness combinations, and compares four agent team communication protocals, and examines single-agent versus multi-agent performance and computational cost.

Critical strengths

First, FATE is grounded in real-world tasks, spans a broad range of domains, includes a substantial number of tasks, includes multimodal inputs, and specifies clearly delineated execution roles for each task. Second, FATE evaluates both final outputs and intermediate artifacts, providing rich information for analyzing and evaluating multi-agent systems. Finally, the paper presents extensive experiments that compare different multi-agent collaboration strategies and assess the effectiveness of multi-agent systems.

Critical weaknesses
  1. FATE’s evaluation metrics rely on both rule-based and LLM-based systems, with the latter primarily answering objective yes/no questions. However, given the complexity and real-world nature of the tasks, models are required to produce multiple deliverable files rather than simple answers to objective questions. It is therefore unclear whether the current metrics adequately capture the quality of these deliverables.
  2. The main experiments reveal several interesting and counterintuitive FE score results, such as GPT-5.6 Sol underperforming GPT-5.6 Terra and Claude Opus 5 underperforming Claude Sonnet 5, but the paper lacks sufficient analysis to explain these findings.
Questions for authors
  1. How were the 300 criterion instances in §2.4 constructed and sampled? Were they drawn from manually constructed valid/invalid deliverables or actual agent outputs, and what was their distribution across domains and positive/negative labels?
  2. Could the authors provide further analysis of these unexpected model rankings, particularly why models with higher ME scores sometimes achieve lower FE scores?
5 comments
AuthorsOpenReview ↗

Response 1

Dear Reviewer GYnt,

Thank you for your careful and insightful review. We appreciate your positive assessment and your recognition of FATE's practical relevance, broad coverage, and comprehensive evaluation. Below, we clarify our evaluation methodology and provide further analysis of the experimental findings. We sincerely hope these responses address your concerns!

W1:FATE's evaluation metrics rely on both rule-based and LLM-based systems, with the latter primarily answering objective yes/no questions. However, given the complexity and real-world nature of the tasks, models are required to produce multiple deliverable files rather than simple answers to objective questions. It is therefore unclear whether the current metrics adequately capture the quality of these deliverables.

We sincerely thank the reviewer for the suggestion. We address the quality of actual deliverables from three perspectives.

(1) FATE follows a common practice among comparable agent benchmarks for evaluating actual deliverables. JobBench [1] and Agents' Last Exam [2] both evaluate actual deliverables using binary judgments for each scoring criterion. FATE follows the same principle by breaking task requirements into verifiable scoring criteria and judging whether the submitted deliverable meets each one.
(2) Binary judgments for individual criteria can assess complex deliverables that span multiple files. For each deliverable, we use the task requirements to design a rubric with multiple specific, verifiable criteria that cover key aspects of quality. Appendix E (Task Examples) illustrates how we apply these criteria to specific deliverables.
(3) Deliverables with higher rubrics scores are preferred by human annotators, showing the current metrics can capture the quality of these deliverables We denote a deliverable pair as a consensus pair if the deliverable with the higher FE score is also preferred by human annotators. Among 200 randomly sampled pairs of outputs from different models on the same task, 91.5% were consensus pairs.

[1]JobBench: Aligning Agent Work With Human Will. Li, Yuetai, et al. (2026)
[2]Agents' Last Exam. Sun, Yiyou, et al. (2026)

W2:The main experiments reveal several interesting and counterintuitive FE score results, such as GPT-5.6 Sol underperforming GPT-5.6 Terra and Claude Opus 5 underperforming Claude Sonnet 5, but the paper lacks sufficient analysis to explain these findings.

We thank the reviewer for pointing out this pattern. As Figure 6 in paper shows, models that perform better on their own do not necessarily collaborate better.
(1) In the setting with a single agent, the results follow an intuitive pattern. GPT-5.6 Sol and Claude Opus 5 outperform GPT-5.6 Terra and Claude Sonnet 5, respectively. GPT-5.6 Sol achieves an FE score of 40.3, compared with 38.8 for GPT-5.6 Terra. Claude Opus 5 achieves an FE score of 28.8, compared with 28.1 for Claude Sonnet 5.
(2) In the setting with agent teams, the FE rankings reverse for both pairs. GPT-5.6 Terra scores 41.1, compared with 36.3 for GPT-5.6 Sol. Claude Sonnet 5 scores 30.1, compared with 29.0 for Claude Opus 5. These results show that models that perform better on their own do not necessarily collaborate better. The collaboration failure modes explained in lines 435 to 448, Figure 9, Table 11, and Appendix E (Error Analysis) suggest directions for future work on improving models' ability to collaborate.

AuthorsOpenReview ↗

Response 2

Q1:How were the 300 criterion instances in §2.4 constructed and sampled? Were they drawn from manually constructed valid/invalid deliverables or actual agent outputs, and what was their distribution across domains and positive/negative labels?

All 300 criterion instances came from actual outputs produced by the eight evaluated agents on benchmark tasks. We randomly sampled from the LLM judge's evaluations of those outputs, with each instance corresponding to one criterion evaluated on one agent output. The sample covered all 14 domains and included 150 positive and 150 negative instances. The domain breakdown was Building (6.33%), Business (11.67%), CFD (6.33%), Education (10.67%), Finance (9.67%), Health (2.67%), Industrial (3.67%), Investigation (3.67%), Legal (9.00%), Media (2.67%), Research (5.33%), Restaurant (9.00%), Social (9.00%), and Tourism (10.33%).

Q2:Could the authors provide further analysis of these unexpected model rankings, particularly why models with higher ME scores sometimes achieve lower FE scores?

Thank you for your insightful question! For some model pairs in our main experiments, the model with the higher ME score has the lower FE score.
(1) ME reflects individual performance, while FE also reflects collaboration. ME evaluates each agent's intermediate outputs, and FE evaluates the team's final deliverables. A model can perform well individually and still struggle to collaborate. Moreover, ME is simpler than FE. As a result, ME yields higher metric values than FE.
(2) Collaboration failures can lower FE even when intermediate outputs are of high quality. We analyze the collaboration failures we observe in our experiments in lines 435 to 448, Figure 9, and Table 11 in paper. Table 11 and Appendix G group these errors into four categories: communication coordination failures, output interface misalignment, state misjudgment, and role boundary violations. These failures help explain why models with higher ME scores sometimes score lower on FE.

AuthorsOpenReview ↗

Response to Reviewer GYnt

Dear GYnt:
Thank you for your insightful feedback! In response to your concerns regarding the quality of deliverable evaluation, the supplementary analysis of experimental phenomena, and the details of the evaluation examples, we have addressed these concerns, and we would like to thank you again for your highly constructive suggestions! The results demonstrate the high quality of the evaluation metrics we employed. We have also provided additional explanations for the relevant experimental phenomena, and incorporated these analyses and details into Appendix A.4, D.3, H.1, H.2 of the revised PDF. If you have any further questions, please feel free to let us know!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for PC and AC

Dear PC and AC,
We sincerely thank you for your time and hard work throughout the review process! We are pleased that all three reviewers give very positive score and suggestions! The reviewers rate the paper 4 (Clear acceptance), 3 (Weak acceptance), and 3 (Weak acceptance), respectively.
(1) We are especially grateful to Reviewer 1vRk for the Clear acceptance and for highlighting our work’s strong practical value for industry and a broad range of users, high-quality data validated by domain experts, and comprehensive evaluation of communication protocols and state-of-the-art agentic harnesses.
(2) We thank for Reviewer 4PxR for their consistently positive score and feedback, and for highlighting FATE’s realistic professional tasks, expert-validated tasks and evaluators, and separate evaluation of individual contributions and final team outcomes.
(3) We thank Reviewer GYnt for their consistently positive score and feedback, and for recognizing FATE’s broad coverage of real-world tasks, clearly defined occupational roles, and extensive experiments comparing collaboration protocols.
To better assist your evaluation, we have prepared a clear and concise summary of the entire discussion.

Strengths Highlighted by the Reviewers

1. Realistic Collaborative Work. FATE focuses on "professional tasks where agents actually need to work together," with role-specific files and permissions, artifact handoffs, and "real materials in their original formats" (Reviewer 4PxR).
2. Broad and Multimodal Coverage. FATE "spans a broad range of domains," includes "a substantial number of tasks" and "multimodal inputs" (Reviewer GYnt).
3. Expert-Led Construction and Quality Control. The benchmark is built through a "rigorous quality control process with the help of domain experts," and its evaluators are tested with valid and invalid outputs (Reviewers 1vRk, 4PxR).
4. Evaluation of Members and Teams. FATE evaluates "both final outputs and intermediate artifacts," giving a fuller picture of team performance than evaluating final outputs alone (Reviewers GYnt, 4PxR).
5. Systematic Experimental Comparisons. The experiments compare "4 different communication protocols" and use "relevant state-of-the-art agentic harnesses," providing a broad assessment of multi-agent collaboration (Reviewers 1vRk, GYnt).

Our Responses and Clarifications

During the rebuttal, all reviewer suggestions concerned clarifications and supplementary details. These clarifications and additions do not affect any of the conclusions or observed phenomena reported in the paper. We have successfully addressed all the reviewers' suggestions. Human and cross-judge validation demonstrates high agreement, supporting the high reliability of our evaluation. We provide further clarification of the FE score computation and the individual and team performance findings, both already presented in the paper. We also clarify experimental details, such as the distinction between the task-subset and full-task-set configurations, and update the corresponding figure.

During the rebuttal, we have successfully addressed all the reviewers' suggestions, we are pleased to see the reviewers confirm in their responses that we have addressed all their suggestions through additional evidence, clarifications, and analyses, which have been incorporated into Appendices A.1, A.4, D.3, H.1, and H.2.

We thank you for overseeing the review process and the reviewers for their careful feedback and thoughtful suggestions. We hope you will consider all our responses in your evaluation, and we thank you again for your time and effort.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.