acceptodds
Under review as a conference paper at ICLR 2027

PowerBench: A Benchmark for Agentic Retrieval and Reasoning in Power Systems

Abstract

Large language model (LLM) agents offer new opportunities for automated analysis in industry. However, rigorous evaluation of such agents — for example, within power system scenarios — remains hindered: real operational data are confidential, and existing public resources fail to fully capture the chained dependencies and heterogeneous evidence. To address this gap, we propose PowerBench, comprising (1) a generation framework that derives interconnected heterogeneous operational data through a common dependency chain, and (2) a synthetic dataset generated by this framework. The dataset covers 761 devices across 100 device types, with 13.35 million hourly telemetry records spanning two years and 24,939 operational documents. Building on this dataset, we construct 300 questions across three task families that evaluate frontier LLMs’ ability to complete analysis tasks that require autonomous evidence retrieval and reasoning across interconnected and heterogeneous data under restricted tool calls and time budgets. Results demonstrate that the evaluated frontier LLMs remain challenged on these tasks: the best model reaches only 74.2% joint accuracy. Our trace analysis further reveals that model performance varies across evidence discovery, content retrieval, tool use, reasoning over evidence, and answer submission. These findings provide detailed insights for evaluating LLM agents and guiding their reliable deployment in industry. The framework, dataset, and benchmark tasks are available at https://anonymous.4open.science/r/PowerBench-9E89/.

Primary area:
datasets and benchmarks
Keywords:
heterogeneous datasets, synthetic data, agentic evaluation, retrieval and reasoning, large language models, power equipment operation and maintenance

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_wzq1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces PowerBench, a benchmark for evaluating LLM-based agents on retrieval and reasoning over heterogeneous power-system data. It provides an LLM-assisted data generation framework and a synthetic dataset containing 761 device instances across 100 device types, 13.35 million hourly telemetry records, and 24,939 operational documents. The benchmark comprises three task families: Temporal Retrieval and Reasoning (TRR), Contextual Retrieval and Reasoning (CRR), and Comparative Retrieval and Reasoning (CoRR). Experiments on nine frontier LLMs evaluate answer accuracy, evidence retrieval, tool reliability, and interaction efficiency under fixed tool-call and time budgets. The results reveal substantial performance gaps, particularly in multi-device comparative reasoning.

Critical strengths
  1. Structured heterogeneous data generation. The proposed framework establishes explicit dependencies among device metadata, operational rules, time-series telemetry, and operational documents. Executable rules and programmatic verification provide a useful mechanism for constructing internally consistent synthetic data with verifiable answers.

  2. Fine-grained agent evaluation and failure attribution. Beyond final-answer accuracy, PowerBench tracks evidence retrieval, tool execution, and answer-generation failures. This enables more informative diagnosis of agent bottlenecks than evaluation based solely on end-task accuracy.

Critical weaknesses

1. The operational scope and semantics of the device-centric tasks remain insufficiently characterized.

Although PowerBench introduces three task families, their operational definitions remain relatively narrow. TRR focuses on analyzing telemetry following an inspection, CRR on applying predefined rules under contextual conditions, and CoRR on comparing metrics across devices. These tasks primarily evaluate evidence retrieval, rule-based computation, and structured aggregation, while the intended notion of interconnected power-system analysis is not sufficiently formalized. In particular, the benchmark does not clearly establish how these tasks represent practical end-to-end equipment operation and maintenance (O&M) workflows, such as online condition monitoring, fault diagnosis, and predictive maintenance. The distinction between cross-source data linkage and genuinely interconnected operational reasoning also remains unclear.

2. The actual task coverage is substantially narrower than the paper's broad positioning suggests.

The paper presents PowerBench as a comprehensive benchmark for agentic reasoning in power systems, but the evaluation is largely restricted to three device-oriented retrieval and analysis workflows. Important capabilities relevant to real-world power-system agents are not explicitly benchmarked, including actionable fault diagnosis, predictive maintenance, grid-level situational awareness, cascading-fault diagnosis, topology-aware analysis, optimal power flow, and decision support for electricity-market bidding. These tasks involve distinct forms of physical constraint reasoning, temporal evolution, cross-component dependency analysis, or decision-making that cannot be adequately represented by the current task families alone. Although the authors acknowledge omitted tasks in Appendix F, the gap between the broad motivation and the narrow operational coverage limits the benchmark's demonstrated contribution to power-system agent research.

3. The importance and unique contribution of time-series reasoning are not adequately demonstrated.

Despite emphasizing heterogeneous text and time-series evidence and providing millions of hourly telemetry records, the benchmark does not sufficiently isolate or characterize the temporal reasoning capabilities it evaluates. The current tasks mainly use telemetry to compute statistics over specified windows, identify event occurrences, and evaluate predefined operational rules, often guided by textual reports. Consequently, it remains unclear whether the benchmark meaningfully evaluates temporal-pattern understanding, evolving equipment conditions, cross-variable temporal dependencies, or event progression beyond relatively straightforward numerical computation. Moreover, the scale of the synthetic telemetry dataset does not itself establish the realism or complexity of its temporal dynamics. Without modality-ablation studies or controlled comparisons against simplified temporal inputs, the paper does not adequately demonstrate how much the time-series modality contributes to agent performance or whether its temporal reasoning claims are justified.

Questions for authors

1. Task definition and operational relevance. How do TRR, CRR, and CoRR systematically map to representative power-equipment O&M workflows? What constitutes interconnected reasoning beyond retrieving linked documents and telemetry or comparing independently analyzed devices? Could the authors clarify the task taxonomy and explain which practical agent capabilities each task is designed to evaluate?

2. Benchmark scope and coverage. Given the stated ambition of evaluating LLM-based agents for power-system analysis, what is the rationale for selecting only the three current task families? How do the authors justify the benchmark's claimed comprehensiveness despite excluding workflows such as online condition monitoring, predictive maintenance, cascading-fault diagnosis, and topology-aware grid analysis? A clearer coverage analysis or evaluation of additional representative tasks would help establish the benchmark's practical significance.

3. Temporal reasoning and modality contribution. What proportion of benchmark questions genuinely requires temporal reasoning, rather than textual evidence retrieval, static metadata lookup, or predefined-rule evaluation? Can the authors provide text-only versus text-plus-time-series ablations, alongside controlled tests using temporally shuffled or aggregated telemetry, to quantify the contribution of temporal information? How do they validate that the generated telemetry captures sufficiently realistic temporal dependencies and operating dynamics?

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer wzq1 for the detailed comments. We address the three questions below.

W1/Q1: Task definitions and operational relevance.

11 Mapping to O&M workflows. The three families are steps inside equipment inspection, fault reassessment, and maintenance prioritization, not the whole of it.

  • TRR follows an inspection: engineers identify readings outside the normal range and track how those metrics move over the next 72 hours.

  • CRR follows a reported event: analysts revisit the associated alarm rule, rescale its threshold by the load class inferred from the narrative, and replay it without resetting the consecutive state that precedes the window.

  • CoRR is the comparison that precedes a maintenance decision, where each unit selects its own metric on its own evidence and the units are then ranked on a common scale.

22 What "interconnected" means. It means that a result derived from one source determines what the agent must retrieve from, or how it must compute on.

  • In TRR the metric to analyse is not given: it must be derived by comparing the inspection readings with the normal mean and standard deviation in the metric catalog, and without that step the agent does not know which series to read.

  • In CRR the threshold is not fixed: it follows from the load category in the fault narrative, and the adjusted rule must then be replayed step by step with consecutive-state carry-over, so the text determines the parameters of the time-series computation.

  • In CoRR, each device may select a different metric with a different unit, and the results become comparable only after standardisation against that device's own catalog statistics, so each chain of report, catalog and telemetry window has to be kept apart before the devices are compared.

Each chain is stored as an executable gold trace over device resolution, metric selection, context branching, window reads and aggregation, and an item is admitted only if the trace replays on the frozen corpus.

33 Taxonomy and capabilities. The progression runs from single-device document-guided analysis TRRTRR, through single-device context-dependent rule execution CRRCRR, to multi-device composition (CoRR), each isolating one capability so that failures can be attributed rather than merely counted. All three also test device resolution from business-level clues among hard negatives, planning under fixed tool and time budgets, and schema-valid structured output. We will add this taxonomy to the main text.

W2/Q2: Benchmark scope and coverage.

11 Why these three. Each family satisfies three criteria: it requires the dependency defined above; its answer can be computed and verified programmatically inside the generated world, so no manual annotation is needed for the gold; and it corresponds to a recurring step in equipment operation and maintenance. Together they span one device with document-guided time-series analysis, one device with context-dependent rule execution, and composition across several devices, which is what makes failure attribution possible.

22 The claim of comprehensiveness. Our wording described the evidence structure: heterogeneous sources linked through a shared dependency chain,not coverage of power-system tasks, and we agree that it reads as the latter. Cascading-fault diagnosis, topology-aware analysis, optimal power flow and market bidding require network models and system-level physical constraints that a device-centric generator does not produce, so they are outside the current scope by design. In the revision we will describe the scope precisely as equipment-level operation-and-maintenance retrieval and reasoning and remove the claim.

33 Coverage. Online condition monitoring and fault diagnosis are partly covered: TRR and CRR analyse monitoring telemetry and alarm rules, but retrospectively rather than in a streaming setting, and CRR does not require root-cause identification. Predictive maintenance is not covered, as the current tasks contain no forecasting targets, and the grid-level workflows above are out of scope for the reason just given. The framework itself extends further than the frozen set: the released code implements fourteen task-family variants, three of which are reported and the rest used in pilot runs. Because every item carries executable gold, new families can be added without annotation. We will include this coverage analysis in the revision and treat predictive-maintenance-style tasks as the priority extension.

AuthorsOpenReview ↗

Rebuttal by Authors [Part 2]

W3/Q3: Temporal reasoning and modality contribution.

11 How much of the benchmark is temporal.

Every question needs telemetry for full credit: part of each answer must be computed from an hourly window, and the leakage audit rejects items that leak gold values or otherwise make the required answer recoverable from the documents the question points to without reading the series. Counting required answer leaves (for CoRR, excluding device aliases), telemetry-dependent fields are half of TRR and CRR and more than 70% of CoRR. Moreover, every question contains at least one field that depends on temporal order rather than on values alone—the first time the maximum occurs TRRTRR, the first trigger time and trigger-hour count under consecutive-hour conditions CRRCRR, or the first peak time per device (CoRR). CRR in particular is a step-by-step rule replay with consecutive-state carry-over across the fault-window boundary, not a static lookup.

22 Ablations.

We will not add text-only, time-shuffled, or aggregated-telemetry runs. Each setting changes the evidence available to the agent, so it requires a modified corpus and a separate evaluation of all nine models. The current evaluation already comprises 2,700 scored runs and about 1.15 billion input tokens, and we will not repeat it under altered inputs.

33 Realistic dynamics.

PowerBench does not aim to reproduce the dynamics of real equipment; real operational data are confidential and hard to obtain, so it builds a synthetic world inspired by power-system settings. For most device types, background series are mean-reverting AR11 paths φ=09˙85φ = 0\.985 clipped inside configuration-aware quiet/safe bands; faults and near-misses are injected with rule-specific patterns, and each event is verified against its target rule, with near-misses backed off toward the background until the rule stays inactive. What we validate is internal consistency, not statistical realism: background does not trigger the screened rules, retained faults trigger theirs, near-misses do not, and gold answers replay on the frozen corpus. We agree that the current tasks exercise windowed statistics, first-occurrence localisation, and stateful replay rather than richer temporal structure such as open-ended forecasting or free-form cross-variable dynamical modelling, and we will say so explicitly. Families built on trends and cross-variable dependencies are a natural extension for a later version.

Reviewer_wzq1OpenReview ↗

Response to Authors

Thanks for your rebuttal. According the narrow contributions and incomplete coverage of this benchmark work, I choose to keep my original score towards rejection.

Reviewer_W8rmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This is a paper for benchmark. The authors introduces PowerBench, a benchmark for evaluating whether agents can retrieve and reason on power-system data. PowerBench combines device metadata, rules, and operational reports. The dataset has 300 questions in 3 task families and evaluates 9 LLMs under fixed tool and budgets. Experiments shows the models performs better on single-device tasks than multiple ones. Also the paper uses tool traces to analyze where agenets fail.

Critical strengths
  1. This paper contributes a new benchmark, it will be useful to support future studies in this domain.

  2. This paper has the failure cases analyze. It's good to study where is the gap.

  3. The executable rules make the benchmark easy to reproduce and audit.

Critical weaknesses
  1. For a benchmark paper, I think the most important thing is the data's quality. This paper includes detailed description of the dataset, but human or expert validation is also essential for this work. But I didn't find the human validation.

  2. The benchmark size is relatively small, it only has 300 questions across 3 task families. Hard to cover all important power-system tasks.

  3. The cost analysis is limited because the tool efficiency is reported as a relative score based on call counts. This paper doesn't have a clear comparison of actual cost or latency across models.

  4. The protocol describes a 300-second retrieval budget and state total latency may exceed it slightly. But Appendix E reports mean latencies around 495 seconds for DeepSeek on CoRR. Please clarify what's the difference and how these values are measured.

Questions for authors

Please see the weaknesses above.

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer W8rm for careful reading and useful comments. We respond to each point below.

W1: Data quality.

We thank the reviewer for raising this point and would like to clarify the design goal of PowerBench:Real-world power operation data are confidential and rarely accessible.PowerBench is therefore not intended to replicate actual operational scenarios or the statistical distributions of real equipment.Instead, it constructs a coherent synthetic world inspired by power-system settings, in which agentic retrieval and reasoning over interconnected heterogeneous evidence can be evaluated in a controlled and verifiable way.

We take several measures to ensure data quality:

Domain-grounded design. The 100-type device taxonomy was manually curated from public industry materials and reviewed by domain practitioners. Device instances, metrics, and rules are generated under this taxonomy with explicit engineering constraints, as described in Appendix A.

Internal consistency by construction. All modalities are derived from a shared dependency chain. Background telemetry is screened against the applicable rules, every retained fault must trigger its target rule, and every near-miss must leave that rule inactive. Documents are rendered from the same structured state, so their quantitative claims can be traced back to the metadata, rules, and telemetry.

Verifiable labels. Every gold answer is recomputed by executable replay over the frozen corpus, and an item is accepted only if the replay matches the stored gold.

Together, these measures ensure that every question is answerable and programmatically verifiable within the synthetic world. This is the property on which the benchmark's measurements rely. We will state this scope more explicitly in the revision. We also agree that a systematic expert assessment of the generated content would further strengthen the benchmark, and we plan to include one in a future version.

W2: Benchmark size and coverage.

We thank the reviewer for this comment. Our choice of 300 questions reflects a deliberate trade-off between coverage and evaluation cost.

Evaluation cost. Each PowerBench question is a multi-step agentic episode rather than a single-turn query. Across the nine evaluated models, a question requires 4–43 tool.The full evaluation comprises 2,700 scored runs and about 1.15B input tokens. Scaling the question set substantially would multiply this cost for every model evaluated on the benchmark, including models evaluated by future users.

ModelNumberMean input tokensMean output tokensMean tool callsMean latency (s)
GPT-5.6-Sol300597,6933,72121.3157.1
DeepSeek-V4-Pro300445,01624,80330.8418.3
Qwen3.8-Max300322,24112,11438.5313.6
Llama-4-Maverick30012,2951,082435.8
Gemini-3.6-Flash3001,111,01525,16831230.1
Doubao-Seed-Evolving300429,34814,41618.4426.3
Kimi-K3300275,6557,40027.5338.4
MiniMax-M3300284,13416,13742.6132.6
GLM-5.3300366,38414,08228250.7
Nine models2,7001.153B (sum)35.7M (sum)4.0–42.6—

Extensibility. PowerBench is a generation framework as well as a fixed question set. The released code already implements 14 task families. The main benchmark freezes three of them, and the others were used in pilot runs AppendixB6˙Appendix B\.6. Because every item is generated with an executable gold answer and verified automatically, new questions and task families can be added without manual annotation. We will expand the benchmark along these lines in future versions.

We also agree that the current task families do not cover every important power-system task. In the revision, we will state this scope more precisely.

AuthorsOpenReview ↗

Rebuttal by Authors [Part 2]

W3: Absolute cost.

We logged the absolute cost of every run but omitted it from the paper. Instead, the table reports recorded token usage and end-to-end latency by model.

ModelNumberMean input tokensMean output tokensMean tool callsMean latency (s)
GPT-5.6-Sol300597,6933,72121.3157.1
DeepSeek-V4-Pro300445,01624,80330.8418.3
Qwen3.8-Max300322,24112,11438.5313.6
Llama-4-Maverick30012,2951,082435.8
Gemini-3.6-Flash3001,111,01525,16831230.1
Doubao-Seed-Evolving300429,34814,41618.4426.3
Kimi-K3300275,6557,40027.5338.4
MiniMax-M3300284,13416,13742.6132.6
GLM-5.3300366,38414,08228250.7
Nine models2,7001.153B (sum)35.7M (sum)4.0–42.6—

Input tokens are accumulated across all interaction turns and therefore include conversation context that is sent repeatedly. TE is retained only as a normalized view of tool-call counts.

W4: Clarification on the latency measurement.

We thank the reviewer for pointing out this inconsistency. The two numbers measure different quantities, and we clarify our implementation below.

The budget. The 300 s budget applies only to the tool-enabled retrieval and reasoning phase. During this phase, every model request is limited to the remaining time, so this phase never exceeds 300s. When the budget is reached, tools are disabled and no further tool call is executed. The model then receives one final no-tool request to submit its answer using only the evidence it has already collected. This final request is not counted toward the 300 s budget.

The reported latency. Latency is measured end-to-end, from the start of a run until the final answer is returned, and therefore includes the final answer submission. The numbers in Appendix E are end-to-end latencies.

Why DeepSeek-V4-Pro reaches about 495 s on CoRR. In 136 of 150 CoRR runs, DeepSeek-V4-Pro reached the 300 s budget. Its final answer submission then took 191.8s on average, because it must combine evidence from 3–5 devices into a single answer and the model produces long outputs.

In the revision, we will correct it and report the retrieval phase and the final-answer phase separately.

Reviewer_xjiBOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This is yet another benchmark dataset of the industrial domain. This paper mainly describes the dataset generation and benchmark.

Critical strengths

The work is an increment, but it makes an interesting point about the need for a dataset for testing an AI Agent for domain-specific tasks.

Critical weaknesses

I read the paper without using any LLM and drafted the review. I used no LLM anywhere in the review process.

  • The main takeaway from the introduction is that the paper aims to provide a benchmark for an industrial domain application, noting a lack of relevant work in the domain targeting a) retrieval and reasoning aspects. I would like to point out a recent, popular work titled AssetOpsBench for the Industrial Operation Domain, which makes a similar claim. Please use these as a starting point, and you will find many relevant papers. As a result, please revise Table 1.

  • My second concern about another important work in the field, TaskBench, is that it can also generate a complex workflow, which could be a good candidate for the Task Complexity bullet point in Table 1. Discussion centered around existing efforts in the industrial domain is relevant and can be swapped with HotPotQA, etc. Such as FailureSensorIQ or IndustryBench. Please conduct a thorough literature review to identify relevant datasets.

  • Another unclear point is whether the task's complexity stems from the long horizon or from what makes the power-industry task different, assuming the related work above is conducted. We need a unique value proposition for these industrial assets.

  • Figure 1 is colorful when printed, and the focus should be on making it readable. Currently, many colors distract from the content. The contribution bullets (especially the 3rd) don't indicate the scale of the experimental work, e.g., 5 frontier models, OSS agents, etc. Citing and building on existing work would strengthen the paper's reusability, but the Introduction currently doesn’t make it clear which parts are novel and which can be built using existing approaches. A justification is needed.

  • Related Work section should include more relevant work from industrial applications. Given that the paper's primary focus is benchmarking and datasets, I expect a more in-depth discussion of relevant papers.

  • Section 3: Given this is a practical application, how does the industrial application define these terms? Suppose we want to represent these assets in an SAP system; how should we present them? The paper needs to go beyond a research article and show how its terminology aligns with existing Products in the domain. Especially Equation 1. For example, why do we need to create a manually curated taxonomy? We already have ISO standards or a Wikipedia article. Clearly, this shows a gap in how papers should define terminology. Please use existing literature to support the design. Section 3 refers to Appendices A.1 to A.6, but a better writeup might include an example dataset in the paper. I did not read the appendix; I expect one example to be the main part of the paper.

  • Section 3.1 needs citations to existing generators; I don't see any new invention or method. Long ago, given a time series T, we injected a fault, like a peak, etc. So what is novel here? What aspect of the power industry is addressed here, and what knowledge is brought here?

  • Section 3.2 did not discuss power-industry-specific metrics in depth. Who provided these metrics, and how were they decided or screened? I reviewed the repository, and it includes some information; however, the main paper needs to provide the details.

  • I do see some overlap of task families in Figure 2: temporal retrieval and reasoning, with the literature. Please search IoT, LLM, etc you will find a lot of work on temporal queries for agents.

  • I think the authors need to conduct an in-depth study of their three task categories; the literature already has plenty of work on time series reasoning and temporal reasoning. The current related work section does not adequately acknowledge these works. Second, how tasks are constructed is not defined properly. Please review the task bench; others define these tasks methodologically. Overall, I read the main section, but it is difficult to identify the key methodological innovation.

  • Harbor is becoming a standard platform for running benchmarks, and I would appreciate it if the authors adopted its framework. This would help establish the authenticity of the modeling approach.

  • Ideally, a benchmark and leaderboard should uncover difficult problems to leave room for future work. In the current setting, frontier models seem well equipped in terms of performance. Maybe a time-to-completion measure in the task study could establish a baseline, or a human study.

  • Overall, the paper is in an interesting direction but needs major work on citing related work and properly analyzing which challenge tasks PowerBench can support.

  • I also reviewed the references, and I think the current format is good, but the work needs more peer-reviewed papers to be technically solid. We also need new, exciting results that reflect the model provided.

Questions for authors

I have understood the paper, and I have no questions that can change my assessment.

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer xjiB for the careful reading and the insightful comments. We respond to each point below.

Related work and Table 1. We did not cover AssetOpsBench, TaskBench, FailureSensorIQ, IndustryBench, or the existing literature on temporal reasoning and temporal queries for agents. We will conduct this literature review and revise Table 1, the Related Work section, and the discussion of Figure 2 accordingly. We will also clarify which difficulty comes from the industrial setting and which comes from the length of the task.

Figure 1 and the contributions. We thank the reviewer for the feedback on Figure 1. We will revise the figure so that the pipeline and contribution claims are easier to read at a glance. We also agree that the contribution statements, especially the third, do not indicate the scale of the experiments. We will state that the evaluation covers nine models, 300 questions, and 2,700 scored runs.

A worked example, Section 3.2, and task construction. Agreed on all three. We will move one complete item into Section 3 and bring the provenance and screening of the metric catalogs and thresholds, currently only in the appendix and the code, into Section 3.2. Task construction belongs in the main text as well: each candidate carries a private gold, is replayed from the released telemetry, rules and documents, and is admitted only if the replay matches, with the wording produced separately from the answer and audited for leakage.

Terminology, Equation 1, and the curated taxonomy. We agree that Equation 1 and the surrounding terminology should be connected to existing industrial standards and asset representations. The 100-type device taxonomy was manually curated from public industry materials and reviewed by domain practitioners. We will use existing literature to support this design and explain how it relates to established classifications.

Harbor. We appreciate this suggestion and will consider providing a Harbor-compatible adapter.

Model performance and time to completion. We respectfully note that the evaluated models are not yet well equipped for these tasks. The best Macro-JA is 74.2%. On CoRR, the best model reaches 48.7% joint accuracy, the second reaches 13.3%, and the other seven models reach at most 2.7%. We already record the end-to-end time to completion for every run. For example, GPT-5.6-Sol takes 157 seconds per question on average, while DeepSeek-V4-Pro takes 418 seconds. We will report these results. A human study would be a valuable addition, and we will consider it in a future version.

We thank the reviewer again for the helpful suggestions.

I acknowledge that I read the author's response. In this rebuttal, the author promised to fulfill all the requests. I maintain my score, as the work still needs improvement. Idealy has potential once implemented.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.