acceptodds
Under review as a conference paper at ICLR 2027

ChatDev 3.0: Enabling Distributed Multi-Agent Collaboration under Execution Asymmetry

Abstract

LLM-based agents are expanding into complex digital workflows that span diverse users and devices with heterogeneous information, authority, and execution states. We characterize these environments by execution asymmetry: while information can cross boundaries through communication, actions remain bounded by user authorization and environment-local interfaces. We present ChatDev 3.0, a distributed multi-agent runtime that explicitly structures collaboration along user-authority and environment-locality boundaries. ChatDev 3.0 employs User Agents to maintain user intent, context, and authority, and Device Agents to execute delegated tasks through local observations and action interfaces. A unified communication protocol and an asynchronous event-driven runtime allow agents to operate concurrently and synchronize only when task dependencies require it. Evaluations on cross-user coordination and cross-device execution benchmarks show that ChatDev 3.0 achieves higher aggregate task success than the evaluated baselines, while also improving execution efficiency in cross-device tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multi agent system, large language model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NYNkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ChatDev 3.0, a distributed multi-agent system designed to coordinate tasks across users and devices while respecting their permissions and local execution constraints. It separates user-level coordination from device-level execution and uses a unified communication protocol and asynchronous runtime to support collaboration and parallel execution. Evaluations on TeamBench, MDCBench-CLI, and MDCBench-GUI show improvements over several baselines in task success and execution efficiency, although the system remains slower than a single-agent baseline in some settings. The results are promising, but the limited GUI evaluation and reliability issues observed in a case study leave questions about the system’s robustness and generalizability.

Critical strengths
  1. Clear problem formulation and practical relevance. The paper identifies execution asymmetry as an important challenge in multi-agent collaboration, highlighting the differences in user permissions, information access, and device-level execution capabilities. This provides a clear motivation for designing a distributed agent architecture.

  2. Well-structured system design. ChatDev 3.0 separates user-level coordination from device-level execution and combines a unified communication protocol with an asynchronous event-driven runtime. This design allows agents to collaborate across users and devices while enabling independent tasks to run concurrently, potentially reducing unnecessary waiting and coordination overhead.

  3. Extensive evaluation across different settings. The paper evaluates the proposed system on cross-user coordination, cross-device command-line tasks, and graphical user interface tasks. The reported results show improvements over several baselines in task success rates and execution efficiency, particularly on MDCBench-CLI. These experiments provide encouraging evidence for the practical value of the proposed architecture.

Critical weaknesses
  1. The framework combines role-based agent decomposition, message-based coordination, and asynchronous execution. However, the paper does not clearly isolate the contribution of each component. Without targeted ablation studies, it is difficult to determine whether the reported gains come from the proposed architecture itself or from the combined effect of these design choices.
  2. Limitations in experimental validation. The GUI evaluation includes only 20 tasks, with one trial per task and a single backbone model, making it difficult to assess the robustness and generalizability of the results. In addition, more ablation studies would help clarify whether the improvements come from the proposed architecture, the asynchronous runtime, or other implementation differences.
  3. Insufficient evidence of end-to-end execution reliability and reproducibility. The case study in Appendix E.2 reveals that the system reports successful task completion even though the restaurant-ranking procedure was not successfully executed or independently verified. This raises concerns about the reliability of its completion claims. Furthermore, the availability of the complete source code and experimental materials could not be confirmed from the provided manuscript, limiting independent reproducibility.
Questions for authors

1.Have the authors conducted ablation studies to isolate the contributions of the boundary-aligned architecture, unified communication protocol, and asynchronous runtime to the reported performance gains? 2.How does the system enforce and verify authorization boundaries during cross-user communication and delegated execution, and has it been evaluated against unauthorized access or information leakage?

3 comments

Thank you for your valuable feedback and constructive comments. We sincerely appreciate your time and effort in reviewing our work. And we hope our response adequately addresses your concerns.

To weakness 1 and question 1:

We thank the reviewer for suggesting concrete comparisons to clarify the contributions of the architecture, communication mechanism, and runtime. These components address distinct aspects of distributed collaboration, and the existing evaluation provides complementary evidence supporting their utility. We will clarify the correspondence between these design choices and the reported evidence below.

  1. Boundary-aligned architecture.

User-level and device-level boundaries serve different purposes. In TeamBench, role-specific information and permissions are prescribed by the benchmark; preserving these boundaries is therefore part of the evaluation setting (section 3.2.1). For device-level organization, the evaluated TopoClaw configuration provides a relevant comparison: a single root agent directly invokes device tools and centrally maintains cross-device context, whereas ChatDev 3.0 delegates device-local state and execution to separate Device Agents. The comparisons align the backbone model, device environments, and available action capabilities (section 3.3.1), with Appendix D.1.3 documenting the functional correspondence between the underlying execution interfaces. ChatDev 3.0 achieves an average Pass@1 of 95.00% versus 90.74% for TopoClaw on MDCBench-CLI (Table 2). On MDCBench-GUI, Pass@1 increases from 25.00% to 75.00%, and Process Score from 44.33% to 79.35% (Table 3). These results support the utility of separating device-local execution and context, particularly in GUI environments.

  1. Autonomous communication.

TeamBench Full provides a relevant comparison for the coordination mechanism. Both methods retain the Planner–Executor–Verifier roles and their permission restrictions, with aligned backbone models, initial environments, and graders (section 3.2.1); Appendix D.1.3 further describes their functionally comparable execution capabilities. Within these conditions, ChatDev 3.0 allows agents to initiate communication according to local needs, rather than following the benchmark’s prescribed interaction flow. Average Solution Pass Rate increases from 53.41% to 57.10%, and Verification Pass Rate from 47.73% to 54.55% (Table 1). Figure 3 additionally documents autonomous reciprocal communication and varying interaction depths. The outcome comparisons and behavioral observations jointly support the utility of autonomous coordination.

  1. Asynchronous runtime.

The asynchronous runtime is an efficiency-oriented design: it reduces blocking and allows independent execution branches to progress concurrently under the existing task dependencies. Its scheduling objective is separate from the task-solving objectives of boundary-aligned execution and autonomous coordination; these contributions are therefore decoupled at the level of design objectives and evaluation criteria. Task effectiveness is evaluated through task-completion metrics, whereas runtime efficiency is evaluated directly through execution time, critical-path efficiency, and observed concurrency. Compared with UFO³, average execution time decreases from 321.85 s to 153.83 s, while critical-path efficiency increases from 41.98% to 53.49% (Table 2). Figure 4(b–c) and Appendix Figure 7 further document greater device concurrency and a larger share of concurrent execution time. Together, these measurements provide evidence directly corresponding to the runtime’s efficiency objective.

Taken together, we believe the existing comparisons and mechanism analyses provide appropriate support for these claims, and hope this clarification addresses the reviewer’s concern about their experimental basis.

To weakness 2:

We acknowledge the limitations of the GUI evaluation in terms of the number of tasks and repeated runs. This is partly due to the substantial execution time and higher api cost required by each GUI task. And we would nevertheless like to clarify its role in the paper: it provides complementary evidence on whether the architecture extends from command-line execution to workflows requiring visual observation and interface interaction. Our conclusions about cross-device execution also draw on the broader MDCBench-CLI experiments, which show consistent performance advantages. We therefore view the GUI results as complementary evidence across interaction modalities, to be considered alongside the CLI evaluation.

And although the evaluation is limited in scale, GUI results in Table 3 shows substantial improvements on both metrics. ChatDev 3.0 achieves a Pass@1 of 75.00%, compared with 25.00% for TopoClaw and 35.00% for UFO³, and a Process Score of 79.35%, compared with 44.33% and 59.65%, respectively.

To weakness 3:

We thank the reviewer for examining the consistency between completion claims and execution evidence. We would like to clarify the distinction between the agent’s final self-report in the example and the paper’s assessment of the recorded trajectory. We do not equate an agent's self-reported completion with verified task completion. Appendix E.2 records the User Agent claiming completion, but immediately states: “This is the User Agent’s final claim; the recorded execution does not establish successful completion of the requested script-based ranking.” The concluding “Verification boundary” further explains that neither the ranking script nor the alternative analysis was confirmed to produce a valid ranking, and that the final text cannot serve as an independently computed result. Its conclusion explicitly states that the case represents “not verified completion of every requirement in the original task.” Thus, we do not present this example as verified end-to-end success. Benchmark evaluation likewise does not directly adopt the agent’s self-reported success: Appendix B.2.3 describes manual assessment based on final environment states, execution traces, and the task-specific scoring rubrics in Table 6. Appendix D.1.3 also distinguishes the evidentiary scope of archived case narratives from that of aggregate benchmark results.

And the supplementary materials, including the evaluation materials and source code, were submitted with the manuscript and are available in the supplementary submission. We hope this clarifies where the reproducibility materials can be accessed.

To question 2:

Device access is scoped by user at the execution layer. Each User Agent accesses Device Agents within its user's authorization scope through the corresponding User Server. Cross-user communication may convey information or collaboration requests, but does not automatically transfer device-control authority; device operations must still be initiated by the receiving side within its own authorization scope. Therefore, cross-user communication does not grant unauthorized access to another user's devices. We note that potential information leakage through agent communication is a separate security concern, which is beyond the scope of this work and remains an important direction for future investigation.

Area_Chair_R7MiOpenReview ↗

Does authors' response address your concerns?

Thank you for your review! Could you check the authors’ response and clarify whether it addresses your original concerns? Many thanks!

AC

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

The paper presents ChatDev 3.0, a distributed multi-agent runtime for tasks spanning multiple users and devices. Its motivation is execution asymmetry: communicating information does not transfer the authorization or device-local capabilities needed to act. The system separates user-level coordination from device-level execution, with explicit reply requirements and an asynchronous event-driven runtime. The authors evaluate the system on TeamBench, MDCBench-CLI, and MDCBench-GUI, reporting aggregate task-success improvements over the evaluated baselines and analyzing communication patterns and cross-device concurrency.

Critical strengths

Well-motivated practical problem. Organizing agents around ownership and authorization boundaries is a sensible design principle. Tasks involving multiple users or devices naturally introduce differences in information, permissions, and local state. The paper identifies a relevant coordination problem that agent systems need to address.

Critical weaknesses
  1. The research contribution is insufficiently distinguished from conventional orchestration. While I agree with the motivation, the paper does not clearly establish what is technically novel about its solution. Separating user-facing coordination from execution agents appears to be a conventional architectural pattern, and dependency-aware communication and asynchronous execution are established ideas. The paper does not adequately identify a limitation of existing approaches that requires its particular design, or explain what new principle emerges from combining these mechanisms. Consequently, I am not convinced that the proposed architecture constitutes a sufficiently distinctive research contribution.

  2. The evaluation does not isolate the mechanisms responsible for the improvements. The end-to-end comparisons leave several plausible explanations unresolved: additional communication, better management of device-local context, or improvements in communication and scheduling. The analyses of device counts and communication rounds are descriptive and do not establish the contribution of the proposed components. Without targeted, controlled ablations, the experiments demonstrate differences between complete systems but provide limited support for the paper’s underlying design rationale. This weakens both the explanatory value of the results and the case for the proposed architecture.

Questions for authors
  1. Contribution clarification: What specific technical mechanism or insight distinguishes ChatDev 3.0 from existing systems that separate user-facing coordination from execution and support dependency-aware asynchronous communication? Please identify the concrete limitation of those approaches that the proposed design addresses.

  2. Evidence clarification: Which experiments isolate the benefits of the user/device decomposition, communication protocol, and asynchronous runtime? Are controlled results available that hold the backbone model, tool capabilities, and resource budgets fixed while varying these components? Such evidence would materially affect my assessment by clarifying whether the gains support the proposed design rationale or arise from other differences between the evaluated systems.

3 comments

Thank you for your valuable feedback and constructive comments. We sincerely appreciate your time and effort in reviewing our work. And we hope our response adequately addresses your concerns.

To question 1:

We thank the reviewer for highlighting the need to clarify the paper’s research contributions. We would like to articulate these contributions in three aspects.

  1. Characterization and formalization of execution asymmetry.

Existing research supports multi-agent collaboration through role or expertise specialization [1] and information exchange [2], which is motivated by distributed ability among agents and distributed information among agents. Building on these foundations, we explicitly characterize the execution constraints in these settings as execution asymmetry, distinguishing user-level authority from device-level environment locality. We formalize tasks that require participation from multiple execution domains, framing the problem as organizing collaboration across these domains while preserving their execution boundaries.

  1. Boundary-aligned architecture and runtime mechanisms for execution asymmetry.

Guided by this problem formulation, we compare ChatDev 3.0 with existing systems that address or approximate this class of problems. UFO³ [3] relies on a centralized ConstellationAgent for task orchestration, while the evaluated TopoClaw[4] configuration uses a single root agent to invoke device tools directly. In contrast, ChatDev 3.0 aligns agent responsibilities with execution boundaries, separating user-level intent, context, and authorization from device-local state and execution. Furthermore, rather than prescribing fixed interaction workflows, agents autonomously determine when and with whom to communicate, while a dependency-aware asynchronous runtime enables independent execution branches to proceed concurrently. Together, these mechanisms provide a unified approach to cross-domain collaboration under execution asymmetry.

  1. Empirical validation across users, devices, and interaction modalities.

We evaluate these designs on TeamBench, MDCBench-CLI, and MDCBench-GUI, comparing against existing frameworks capable of handling the corresponding tasks. On MDCBench-CLI, ChatDev 3.0 achieves an average Pass@1 of 95.00%, compared with 90.74% for TopoClaw and 81.30% for UFO³. On MDCBench-GUI, the corresponding results are 75.00%, 25.00%, and 35.00%, respectively (Tables 2–3). Compared with TeamBench Full, average Solution Pass Rate increases from 53.41% to 57.10%, and Verification Pass Rate from 47.73% to 54.55% (Table 1).

We will further emphasize these three contributions in the revised manuscript. We hope this clarification makes their relationship to existing capabilities clearer.

References

[1] D. X. Long et al., “Multi-expert Prompting Improves Reliability, Safety and Usefulness of Large Language Models,” EMNLP, 2024.

[2] W. Liu et al., “Autonomous Agents for Collaborative Task under Information Asymmetry,” NeurIPS, 2024.

[3] C. Zhang et al., “UFO³: Weaving the Digital Agent Galaxy,” arXiv:2511.11332, 2025.

[4] H. Huang et al., “TopoClaw: A Human-Centric and Topology-Aware Agent Operating System,” arXiv:2605.15556, 2026.

To question 2:

We thank the reviewer for emphasizing the importance of distinguishing system-level comparisons, observations of mechanism behavior, and causal attribution. We agree that these forms of evidence should be interpreted carefully. Our position is that the existing experiments provide complementary support for the utility of the proposed designs within the evaluated settings,.

  1. Boundary-aligned architecture.

User-level and device-level boundaries serve different purposes. In TeamBench, role-specific information and permissions are prescribed by the benchmark; preserving these boundaries is therefore part of the evaluation setting (section 3.2.1). For device-level organization, the evaluated TopoClaw configuration provides a relevant comparison: a single root agent directly invokes device tools and centrally maintains cross-device context, whereas ChatDev 3.0 delegates device-local state and execution to separate Device Agents. The comparisons align the backbone model, device environments, and available action capabilities (section 3.3.1), with Appendix D.1.3 documenting the functional correspondence between the underlying execution interfaces. ChatDev 3.0 achieves an average Pass@1 of 95.00% versus 90.74% for TopoClaw on MDCBench-CLI (Table 2). On MDCBench-GUI, Pass@1 increases from 25.00% to 75.00%, and Process Score from 44.33% to 79.35% (Table 3). These results support the utility of separating device-local execution and context, particularly in GUI environments.

  1. Autonomous communication.

TeamBench Full provides a relevant comparison for the coordination mechanism. Both methods retain the Planner–Executor–Verifier roles and their permission restrictions, with aligned backbone models, initial environments, and graders (section 3.2.1); Appendix D.1.3 further describes their functionally comparable execution capabilities. Within these conditions, ChatDev 3.0 allows agents to initiate communication according to local needs, rather than following the benchmark’s prescribed interaction flow. Average Solution Pass Rate increases from 53.41% to 57.10%, and Verification Pass Rate from 47.73% to 54.55% (Table 1). Figure 3 additionally documents autonomous reciprocal communication and varying interaction depths. The outcome comparisons and behavioral observations jointly support the utility of autonomous coordination.

  1. Asynchronous runtime.

The asynchronous runtime is an efficiency-oriented design: it reduces blocking and allows independent execution branches to progress concurrently under the existing task dependencies. Its scheduling objective is separate from the task-solving objectives of boundary-aligned execution and autonomous coordination; these contributions are therefore decoupled at the level of design objectives and evaluation criteria. Task effectiveness is evaluated through task-completion metrics, whereas runtime efficiency is evaluated directly through execution time, critical-path efficiency, and observed concurrency. Compared with UFO³, average execution time decreases from 321.85 s to 153.83 s, while critical-path efficiency increases from 41.98% to 53.49% (Table 2). Figure 4(b–c) and Appendix Figure 7 further document greater device concurrency and a larger share of concurrent execution time. Together, these measurements provide evidence directly corresponding to the runtime’s efficiency objective.

Taken together, we believe the existing comparisons and mechanism analyses provide appropriate support for these claims, and hope this clarification addresses the reviewer’s concern about their experimental basis.

Area_Chair_R7MiOpenReview ↗

Does authors' response address your concerns?

Thank you for your review! Could you check the authors’ response and clarify whether it addresses your original concerns? Many thanks!

AC

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

The paper presents a practical multi agent framework to operate in scenarios which require cooperation among different users and devices. The framework contains three main components: (1) a boundary-aligned distinction into two agents, Device Agents and User Agents; (2) a unified communication protocol that enables autonomous coordination between agents; and (3) an asynchronous event-driven runtime that allows agents to execute tasks concurrently and synchronize when required.

The framework is evaluated on three Benchmarks, namely, TeamBench, MDCBench-CLI and MDCBench-GUI. The results show improvements in task success rates over the evaluated baselines, along with reduced execution time. Overall, the paper aims to provide a unified architecture for multi-agent collaboration across user and device boundaries.

Critical strengths

The motivation of the paper and the problem it targets are sound and well formulated. The distinction between user-level authority and device-level execution is sensible, and the proposed solution is a useful addition to existing multi-agent systems.

The authors also evaluate the framework across multiple benchmarks and backbone models, with results showing improvements over the evaluated baselines, particularly for cross-device execution tasks.

Critical weaknesses

While the problem is well formulated and the proposed solution is sensible, I am not convinced that it offers a substantial improvement over existing multi-agent architectures. Most of the proposed components, including the separation of User and Device Agents, the communication protocol, and the asynchronous execution runtime, build on already established ideas. Existing protocols such as Google's Agent2Agent (A2A) already address several aspects of communication, delegation, and coordination between agents. The paper does not clearly establish what the proposed framework adds beyond integrating these existing ideas. The distinction into User Agents and Device Agents is a clear addition, but I think the authors should also establish how their work differs from existing multi-agent communication protocols.

While the results show improvements over the evaluated baselines, I think a comparison with existing approaches such as A2A would have been particularly useful. Such protocols could also be used to build systems addressing similar execution and coordination problems, and it is unclear whether the proposed framework offers any significant advantage over them. Without such comparisons, it is difficult to assess how much the proposed architecture actually improves upon existing approaches. I also have concerns about the lack of ablation studies, particularly around the three main contributions of the paper. For instance, comparing the proposed architecture against a centralized coordinator with the same execution capabilities would help establish the benefit of separating User Agents and Device Agents. Similarly, evaluating the system with fixed communication patterns instead of autonomous coordination, and with synchronous execution instead of the proposed asynchronous runtime, would help establish how much each component contributes to the overall improvements.

As a minor concern that does not affect my overall assessment, I found parts of the paper difficult to follow, particularly the connection between the formal problem formulation and the proposed solution. Some additional clarity on the individual contributions would be helpful.

Questions for authors

I am willing to increase my score based on the authors' clarifications to the following questions.

  1. Can the authors clarify how the proposed framework differs from existing multi-agent communication protocols such as Google's Agent2Agent (A2A)? In particular, what capabilities or advantages does ChatDev 3.0 provide that could not be achieved by combining existing communication protocols with an appropriate execution runtime. I understand that implementing and running an A2A-based baseline end to end may not be feasible during the rebuttal period, so a clear technical comparison explaining the differences and advantages of the proposed approach would be sufficient.

  2. Can the authors clarify the contributions of the individual components, particularly the separation of User and Device Agents, autonomous communication, and asynchronous execution? For instance, how would the system compare against a centralized coordinator with the same execution capabilities, fixed communication patterns instead of autonomous coordination, or synchronous execution instead of the proposed asynchronous runtime? Ablation results, if available, would be helpful in understanding which components are responsible for the reported improvements.

2 comments

Thank you for your valuable feedback and constructive comments. We sincerely appreciate your time and effort in reviewing our work. And we hope our response adequately addresses your concerns.

To question 1:

We thank the reviewer for highlighting the need to clarify the relationship between ChatDev 3.0 and A2A.

A2A is an interoperability protocol for agents, whereas ChatDev 3.0 is a distributed execution runtime for tasks spanning users and devices. The A2A specification defines capability discovery, task interactions, status updates, and associated interfaces, but does not prescribe how a system should organize its agents, maintain execution contexts, or schedule device operations [1].

We therefore agree that an A2A-based system could be built to address the execution asymmetry studied in this paper. Such a system would still require concrete choices and designs about its architecture, coordination policies, and runtime implementation. Its task performance would depend on these choices rather than on the protocol alone. Our work investigates these system-level contributions and evaluates their effectiveness in cross-user and cross-device settings.

[1] A2A Protocol Specification, v0.3.0. https://a2a-protocol.org/v0.3.0/specification/

To question 2:

We thank the reviewer for suggesting concrete comparisons to clarify the contributions of the architecture, communication mechanism, and runtime. These components address distinct aspects of distributed collaboration, and the existing evaluation provides complementary evidence supporting their utility. We clarify the correspondence between these design choices and the reported evidence below.

  1. Boundary-aligned architecture. User-level and device-level boundaries serve different purposes. In TeamBench, role-specific information and permissions are prescribed by the benchmark; preserving these boundaries is therefore part of the evaluation setting (section 3.2.1). For device-level organization, the evaluated TopoClaw configuration provides a relevant comparison: a single root agent directly invokes device tools and centrally maintains cross-device context, whereas ChatDev 3.0 delegates device-local state and execution to separate Device Agents. The comparisons align the backbone model, device environments, and available action capabilities (section 3.3.1), with Appendix D.1.3 documenting the functional correspondence between the underlying execution interfaces. ChatDev 3.0 achieves an average Pass@1 of 95.00% versus 90.74% for TopoClaw on MDCBench-CLI (Table 2). On MDCBench-GUI, Pass@1 increases from 25.00% to 75.00%, and Process Score from 44.33% to 79.35% (Table 3). These results support the utility of separating device-local execution and context, particularly in GUI environments.

  2. Autonomous communication. TeamBench Full provides a relevant comparison for the coordination mechanism. Both methods retain the Planner–Executor–Verifier roles and their permission restrictions, with aligned backbone models, initial environments, and graders (section 3.2.1); Appendix D.1.3 further describes their functionally comparable execution capabilities. Within these conditions, ChatDev 3.0 allows agents to initiate communication according to local needs, rather than following the benchmark’s prescribed interaction flow. Average Solution Pass Rate increases from 53.41% to 57.10%, and Verification Pass Rate from 47.73% to 54.55% (Table 1). Figure 3 additionally documents autonomous reciprocal communication and varying interaction depths. The outcome comparisons and behavioral observations jointly support the utility of autonomous coordination.

  3. Asynchronous runtime. The asynchronous runtime is an efficiency-oriented design: it reduces blocking and allows independent execution branches to progress concurrently under the existing task dependencies. Its scheduling objective is separate from the task-solving objectives of boundary-aligned execution and autonomous coordination; these contributions are therefore decoupled at the level of design objectives and evaluation criteria. Task effectiveness is evaluated through task-completion metrics, whereas runtime efficiency is evaluated directly through execution time, critical-path efficiency, and observed concurrency. Compared with UFO³, average execution time decreases from 321.85 s to 153.83 s, while critical-path efficiency increases from 41.98% to 53.49% (Table 2). Figure 4(b–c) and Appendix Figure 7 further document greater device concurrency and a larger share of concurrent execution time. Together, these measurements provide evidence directly corresponding to the runtime’s efficiency objective.

Taken together, we believe the existing comparisons and mechanism analyses provide appropriate support for these claims, and hope this clarification addresses the reviewer’s concern about their experimental basis.

Area_Chair_R7MiOpenReview ↗

Does authors' response address your concerns?

Thank you for your review! Could you check the authors’ response and clarify whether it addresses your original concerns? Many thanks!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.