arxiv: 2604.22820 · v1 · submitted 2026-04-17 · 💻 cs.MA · cs.AI

Recognition: unknown

Complete Cyclic Subtask Graphs for Tool-Using LLM Agents: Flexibility, Cost, and Bottlenecks in Multi-Agent Workflows

Luay Gharzeddine , Samer Saab Jr

Authors on Pith no claims yet

Pith reviewed 2026-05-10 07:01 UTC · model grok-4.3

classification 💻 cs.MA cs.AI

keywords cyclic subtask graphsmulti-agent workflowsLLM agentstool userevisitationcoordination costtask bottlenecksrecovery exploration

0 comments

The pith

Complete cyclic subtask graphs show when revisitation helps recovery in multi-agent LLM workflows versus when it adds cost or leaves bottlenecks dominant.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes complete cyclic subtask graphs as a testbed for multi-agent tool-using LLM agents that allows unrestricted revisitation of subtasks. By fully connecting subtask nodes and using one router agent to choose transitions in natural language, the setup makes the benefits and costs of flexibility directly measurable. Experiments across TextCraft, ALFWorld, and Finance-Agent identify three regimes: recovery gains from revisits in ALFWorld, efficiency of forward-only in TextCraft, and external bottlenecks in Finance-Agent. This matters because it clarifies when investing in more complex multi-agent coordination is worthwhile compared to simpler single-agent approaches like ReAct.

Core claim

By constructing complete cyclic subtask graphs in which every subtask node connects to all others and a unified state-analysis-and-routing agent selects the next transition according to natural-language criteria, the work demonstrates that this maximal flexibility supports recovery and exploration in ALFWorld, favors simpler forward execution in the prerequisite-chain setting of TextCraft, and fails to overcome retrieval and grounding limits in Finance-Agent, while consistently incurring higher inference costs than a single ReAct agent.

What carries the argument

The complete cyclic subtask graph, a fully connected architecture of executable subtask nodes with a unified routing agent that picks transitions via natural language descriptions.

If this is right

ALFWorld performance improves with the ability to revisit subtasks for recovery and exploration.
TextCraft tasks perform better or equally with restricted forward execution due to lower coordination overhead.
Finance-Agent tasks remain limited by retrieval, grounding, and evidence synthesis regardless of workflow flexibility.
The added flexibility in cyclic graphs leads to substantially higher token usage than a single ReAct agent.
Ablations show that agent strengths, tool specialization, and perturbations affect the observed regimes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Task designers could first test a domain on cyclic graphs to decide whether to invest in revisitation mechanisms or focus on improving base capabilities.
Optimizing the routing agent to avoid unnecessary cycles might reduce the cost penalty while retaining recovery benefits.
The three regimes suggest that no single multi-agent architecture suits all tool-using tasks, pointing toward adaptive or hybrid designs.
Extending the lens to measure specific costs of coordination versus execution could guide more efficient agent orchestration.

Load-bearing premise

A unified state-analysis-and-routing agent can effectively select transitions using natural-language criteria across domains and agent strengths without introducing unmanageable coordination overhead.

What would settle it

If the cyclic graphs showed no performance improvement in ALFWorld over acyclic versions or no cost increase in TextCraft, the claim that they distinguish distinct regimes of benefit, cost, and bottlenecks would be falsified.

Figures

Figures reproduced from arXiv: 2604.22820 by Luay Gharzeddine, Samer Saab Jr.

**Figure 2.** Figure 2: Comparison of graph construction regimes across benchmarks ( [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

**Figure 3.** Figure 3: Success-only subtask visitation decomposition. [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

**Figure 4.** Figure 4: ALFWorld environment-level exploration study: difference in unique [PITH_FULL_IMAGE:figures/full_fig_p028_4.png] view at source ↗

**Figure 5.** Figure 5: Illustrative Spec-Cyc ALFWorld success showing error correction through [PITH_FULL_IMAGE:figures/full_fig_p030_5.png] view at source ↗

**Figure 6.** Figure 6: Illustrative Spec-Cyc ALFWorld failure trace showing oscillation between [PITH_FULL_IMAGE:figures/full_fig_p031_6.png] view at source ↗

**Figure 7.** Figure 7: Illustrative Gen-Cyc TextCraft-3 failure under a specialized raw-materials [PITH_FULL_IMAGE:figures/full_fig_p032_7.png] view at source ↗

**Figure 8.** Figure 8: Illustrative Gen-Cyc TextCraft-3 failure under a generalist executor. The [PITH_FULL_IMAGE:figures/full_fig_p033_8.png] view at source ↗

read the original abstract

Long-horizon tool-using tasks sometimes benefit from revisiting earlier subtasks for recovery and exploration, but added multi-agent workflow flexibility can also introduce coordination overhead and substantial inference cost. We study complete cyclic subtask graphs, a deliberately maximally flexible multi-agent architecture in which executable subtask nodes are fully connected and a unified state-analysis-and-routing agent selects transitions using natural-language criteria. This makes unrestricted revisitation explicit and directly analyzable at the subtask level. We evaluate task-specific (Spec-Cyc) and benchmark-generic (Gen-Cyc) graphs on TextCraft, ALFWorld, and Finance-Agent, with ablations over planner/executor/router strength, tool exposure (generalist vs specialized), $n$-shot successful trajectory summaries, and fault-injected random subtask perturbations. The benchmarks expose three distinct regimes. ALFWorld highlights a setting where explicit revisitation supports recovery and exploration; TextCraft, a largely prerequisite-chain domain, often favors the efficiency of simpler forward execution; and Finance-Agent remains bottlenecked by retrieval, grounding, and evidence synthesis more than by workflow flexibility alone. Shared-win token comparisons further show that the added flexibility can be substantially more expensive than a single ReAct agent. Overall, we use complete cyclic subtask graphs as a maximally flexible experimental lens for measuring when multi-agent revisitation helps, when it mainly adds coordination cost, and when external task bottlenecks dominate.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Complete cyclic subtask graphs work as a clean experimental lens for when multi-agent revisitation helps versus adds cost in LLM tool use, with benchmark-specific regimes and clear token overhead data.

read the letter

The paper's main point is that complete cyclic subtask graphs, with every subtask node fully connected and a single natural-language router handling transitions, give a deliberate maximal-flexibility setup for measuring revisitation effects in tool-using LLM agents. They test both task-specific and generic versions on ALFWorld, TextCraft, and Finance-Agent, running ablations on router strength, tool specialization, n-shot summaries, and fault injection. The results separate three regimes without forcing one story: ALFWorld shows recovery and exploration gains from going back, TextCraft favors simpler forward chains, and Finance-Agent stays limited by retrieval and synthesis rather than workflow structure. Shared token-cost comparisons to ReAct baselines make the coordination overhead concrete and quantifiable. This is straightforward empirical work that isolates the variable they care about. The design avoids circular claims by letting benchmark outcomes drive the conclusions, and the ablations are systematic enough to support the distinctions they draw. A minor soft spot is the consistently high token cost, which the paper treats as a feature of the lens rather than a flaw, but it does limit how far the architecture itself would scale in practice. The router assumption holds in their runs, though it could be brittle with weaker models or noisier state descriptions. The three benchmarks provide useful contrast but are still a narrow slice of possible tasks. This is for people working on multi-agent LLM workflows who need data on flexibility tradeoffs rather than a new production recipe. It deserves a serious referee because the methods are explicit, the metrics are shared, and the questions map directly to design decisions others are making.

Referee Report

0 major / 2 minor

Summary. The paper introduces complete cyclic subtask graphs as a deliberately maximal-flexibility multi-agent architecture for tool-using LLM agents. Subtask nodes are fully connected, and a unified state-analysis-and-routing agent selects transitions via natural-language criteria. Evaluations on TextCraft, ALFWorld, and Finance-Agent include ablations on planner/executor/router strength, tool specialization, n-shot summaries, and fault injection. The work identifies three regimes (recovery/exploration in ALFWorld, forward efficiency in TextCraft, retrieval bottlenecks in Finance-Agent) and quantifies added costs versus ReAct baselines via shared-win token comparisons.

Significance. If the empirical results hold, the complete cyclic subtask graph serves as a valuable experimental lens for isolating when multi-agent revisitation aids recovery versus when it primarily increases coordination cost or when external bottlenecks dominate. Strengths include the systematic ablations across router/tool variants and the cross-benchmark comparison that surfaces domain-specific trade-offs. The shared-win token analysis adds practical relevance for cost-aware workflow design.

minor comments (2)

[Evaluation] The precise definition and computation of 'shared-win token costs' versus ReAct baselines should be stated explicitly (e.g., in the evaluation metrics subsection) to ensure reproducibility across readers.
[Methods] Figure captions or the methods section could include one concrete example of a natural-language transition decision made by the unified router to illustrate the selection criteria.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive and accurate summary of our manuscript on complete cyclic subtask graphs, as well as for recognizing the value of the systematic ablations, cross-benchmark comparisons, and shared-win token analysis. We appreciate the recommendation for minor revision.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper is an empirical evaluation study that defines complete cyclic subtask graphs as a maximally flexible architecture, implements it with full connectivity and a unified LLM router, and measures outcomes via ablations and benchmark comparisons on TextCraft, ALFWorld, and Finance-Agent. No derivations, equations, fitted parameters presented as predictions, or self-citations appear in the provided text or abstract. The three regimes and cost analyses emerge directly from experimental results rather than reducing to input definitions or prior author work by construction. The central claim functions as an experimental lens grounded in external benchmarks and baselines like ReAct.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach assumes the effectiveness of the routing mechanism and that the benchmarks capture relevant aspects of real-world tool-using tasks.

axioms (1)

domain assumption Natural language criteria can be used effectively by the routing agent to select subtask transitions
This is central to the architecture described in the abstract.

pith-pipeline@v0.9.0 · 5562 in / 1198 out tokens · 70974 ms · 2026-05-10T07:01:37.807104+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

40 extracted references · 14 canonical work pages · 1 internal anchor

[1]

In: The Eleventh International Conference on Learning Representations (2022)

Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K.R., Cao, Y.: React: Synergizing reasoning and acting in language models. In: The Eleventh International Conference on Learning Representations (2022)

2022
[2]

In: Proceedings of the 42nd International Conference on Machine Learning (2025)

Erdogan, L.E., Lee, N., Kim, S., Moon, S., Furuta, H., Anumanchipalli, G., Keutzer, K., Gholami, A.: Plan-and-act: Improving planning of agents for long-horizon tasks. In: Proceedings of the 42nd International Conference on Machine Learning (2025)

2025
[3]

In: Conference on Language Modeling (COLM 2024) (2024)

Wu, Y., Yue, T., Zhang, S., Wang, C., Wu, Q.: Stateflow: Enhancing LLM task-solving through state-driven workflows. In: Conference on Language Modeling (COLM 2024) (2024)

2024
[4]

In: Forty-first International Conference on Machine Learning (2024)

Zhuge, M., Wang, W., Kirsch, L., Faccio, F., Khizbullin, D., Schmidhuber, J.: Gptswarm: Language agents as optimizable graphs. In: Forty-first International Conference on Machine Learning (2024)

2024
[5]

In: International Conference on Machine Learning (ICML) (2025)

Wang, Z.Z., Mao, J., Fried, D., Neubig, G.: Agent workflow memory. In: International Conference on Machine Learning (ICML) (2025)

2025
[6]

Towards a Science of Scaling Agent Systems

Kim, Y., Gu, K., Park, C., Park, C., Schmidgall, S., Heydari, A.A., Yan, Y., Zhang, Z., Zhuang, Y., Malhotra, M., et al.: Towards a science of scaling agent systems. arXiv preprint arXiv:2512.08296 (2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[7]

Advances in Neural Information Processing Systems 36, 8634–8652 (2023)

Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., Yao, S.: Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems 36, 8634–8652 (2023)

2023
[8]

In: Forty-first International Conference on Machine Learning (2024)

Wang, X., Chen, Y., Yuan, L., Zhang, Y., Li, Y., Peng, H., Ji, H.: Executable code actions elicit better llm agents. In: Forty-first International Conference on Machine Learning (2024)

2024
[9]

In: Findings of the Association for Computational Linguistics: NAACL 2024, pp

Prasad, A., Koller, A., Hartmann, M., Clark, P., Sabharwal, A., Bansal, M., Khot, T.: Adapt: As-needed decomposition and planning with language models. In: Findings of the Association for Computational Linguistics: NAACL 2024, pp. 4226–4252 (2024)

2024
[10]

In: Kwok, J

Coles, A., Karpas, E., Shimony, E., Shperberg, S., Ruml, W.: Concurrent planning and execution using dispatch-dependent values. In: Kwok, J. (ed.) Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25, pp. 8483–8490 (2025)

2025
[11]

In: Alechina, N., Dignum, V., Dastani, M., Sichman, J.S

Shi, H., Sun, Z., Yuan, X., Côté, M.-A., Liu, B.: OPEx: A large language model-powered framework for embodied instruction following. In: Alechina, N., Dignum, V., Dastani, M., Sichman, J.S. (eds.) Proceedings of the 23rd 34 International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2024), pp. 2465–2467 (2024). Extended Abstract

2024
[12]

Re- woo: Decoupling reasoning from observations for ef- ficient augmented language models

Xu, B., Peng, Z., Lei, B., Mukherjee, S., Liu, Y., Xu, D.: Rewoo: Decoupling reasoning from observations for eﬀicient augmented language models. arXiv preprint arXiv:2305.18323 (2023)

work page arXiv 2023
[13]

Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, and 1 others

Shi, Y., Wang, M., Cao, Y., Lai, H., Lan, J., Han, X., Wang, Y., Geng, J., Li, Z., Xia, Z., et al.: Aime: Towards fully-autonomous multi-agent framework. arXiv preprint arXiv:2507.11988 (2025)

work page arXiv 2025
[14]

arXiv preprint arXiv:2506.12508 , year =

Zhang, W., Cui, C., Zhao, Y., Liu, Y., An, B.: Agentorchestra: A hierarchical multi-agent framework for general-purpose task solving. arXiv preprint arXiv:2506.12508 (2025)

work page arXiv 2025
[15]

In: Proceedings of the 42nd International Conference on Machine Learning

Zhang, Y., Liu, X., Xiao, C.: MetaAgent: Automatically constructing multi-agent systems based on finite state machines. In: Proceedings of the 42nd International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 267, pp. 75667–75694 (2025). ICML 2025 Poster

2025
[16]

Autonomous Agents and Multi-Agent Systems 16, 151–185 (2008)

Sims, M., Corkill, D., Lesser, V.: Automated organization design for multi-agent systems. Autonomous Agents and Multi-Agent Systems 16, 151–185 (2008)

2008
[17]

Autonomous Agents and Multi-Agent Systems 25, 46–86 (2012)

Weerdt, M.M., Zhang, Y., Klos, T.: Multiagent task allocation in social networks. Autonomous Agents and Multi-Agent Systems 25, 46–86 (2012)

2012
[18]

Autonomous Agents and Multi-Agent Systems 18, 267–294 (2009)

Jonge, F., Roos, N., Witteveen, C.: Primary and secondary diagnosis of multi-agent plan execution. Autonomous Agents and Multi-Agent Systems 18, 267–294 (2009)

2009
[19]

Gradientsys: A multi-agent llm scheduler with react orchestration,

Song, X., Wang, Z., Wu, S., Shi, T., Ai, L.: Gradientsys: A multi-agent llm scheduler with react orchestration. arXiv preprint arXiv:2507.06520 (2025)

work page arXiv 2025
[20]

G-designer: Architecting multi-agent communication topologies via graph neural networks.arXiv preprint arXiv:2410.11782, 2024

Zhang, G., Yue, Y., Sun, X., Wan, G., Yu, M., Fang, J., Wang, K., Chen, T., Cheng, D.: G-designer: Architecting multi-agent communication topologies via graph neural networks. arXiv preprint arXiv:2410.11782 (2024)

work page arXiv 2024
[21]

In: Proceedings of the 42nd International Conference on Machine Learning

Zhang, G., Niu, L., Fang, J., Wang, K., Bai, L., Wang, X.: Multi-agent architecture search via agentic supernet. In: Proceedings of the 42nd International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 267, pp. 75834–75852 (2025)

2025
[22]

arXiv preprint arXiv:2503.03686 , year=

Ye, R., Tang, S., Ge, R., Du, Y., Yin, Z., Chen, S., Shao, J.: Mas-gpt: Training llms to build llm-based multi-agent systems. arXiv preprint arXiv:2503.03686 (2025)

work page arXiv 2025
[23]

Scaling large-language-model-based multi-agent collaboration

Qian, C., Xie, Z., Wang, Y., Liu, W., Zhu, K., Xia, H., Dang, Y., Du, Z., Chen, 35 W., Yang, C., et al.: Scaling large language model-based multi-agent collaboration. arXiv preprint arXiv:2406.07155 (2024)

work page arXiv 2024
[24]

In: Proceedings of the 41st International Conference on Machine Learning

Smit, A.P., Grinsztajn, N., Duckworth, P., Barrett, T.D., Pretorius, A.: Should we be going MAD? A look at multi-agent debate strategies for LLMs. In: Proceedings of the 41st International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 235, pp. 45883–45905. PMLR, ??? (2024)

2024
[25]

In: Proceedings of the 41st International Conference on Machine Learning

Chen, J.C.-Y., Saha, S., Stengel-Eskin, E., Bansal, M.: MAGDi: Structured distillation of multi-agent interaction graphs improves reasoning in smaller language models. In: Proceedings of the 41st International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 235, pp. 7220–7235 (2024)

2024
[26]

arXiv preprint arXiv:2405.06691 (2024)

Klein, L., Potamitis, N., Aydin, R., West, R., Gulcehre, C., Arora, A.: Fleet of agents: Coordinated problem solving with large language models. arXiv preprint arXiv:2405.06691 (2024)

work page arXiv 2024
[27]

In: The Twelfth International Conference on Learning Representations (ICLR 2024) (2024)

Qin, Y., Liang, S., Ye, Y., Zhu, K., Yan, L., Lu, Y., Lin, Y., Cong, X., Tang, X., Qian, B., Zhao, S., Hong, L., Tian, R., Xie, R., Zhou, J., Gerstein, M., Li, D., Liu, Z., Sun, M.: Toolllm: Facilitating large language models to master 16000+ real-world apis. In: The Twelfth International Conference on Learning Representations (ICLR 2024) (2024)

2024
[28]

Retrieval models aren’t tool-savvy: Benchmarking tool retrieval for large language models.arXiv preprint arXiv:2503.01763, 2025

Shi, Z., Wang, Y., Yan, L., Ren, P., Wang, S., Yin, D., Ren, Z.: Retrieval models aren’t tool-savvy: Benchmarking tool retrieval for large language models. arXiv preprint arXiv:2503.01763 (2025)

work page arXiv 2025
[29]

In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp

Qin, S., Zhu, Y., Mu, L., Zhang, S., Zhang, X.: Meta-tool: Unleash open-world function calling capabilities of general-purpose large language models. In: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 30653–30677 (2025)

2025
[30]

Advances in Neural Information Processing Systems 37, 126544–126565 (2024)

Patil, S.G., Zhang, T., Wang, X., Gonzalez, J.E.: Gorilla: Large language model connected with massive apis. Advances in Neural Information Processing Systems 37, 126544–126565 (2024)

2024
[31]

arXiv preprint arXiv:2508.08882 (2025)

Wang, D., Yang, J., Li, W., Liang, J., Li, Y.: Reducing cognitive overhead in tool use via multi-small-agent reinforcement learning. arXiv preprint arXiv:2508.08882 (2025)

work page arXiv 2025
[32]

MiroMind AI Team

Song, L., Liu, J., Zhang, J., Zhang, S., Luo, A., Wang, S., Wu, Q., Wang, C.: Adaptive in-conversation team building for language model agents. arXiv preprint arXiv:2405.19425 (2024)

work page arXiv 2024
[33]

36 Autonomous Agents and Multi-Agent Systems 38 (2024)

Yang, M., Zhao, K., Wang, Y., Dong, R., Du, Y., Liu, F., Zhou, M., U, L.H.: Team-wise effective communication in multi-agent reinforcement learning. 36 Autonomous Agents and Multi-Agent Systems 38 (2024). Article 36

2024
[34]

In: Proceedings of the 42nd International Conference on Machine Learning (2025)

Huang, J.-t., Zhou, J., Jin, T., Zhou, X., Chen, Z., Wang, W., Yuan, Y., Lyu, M.R., Sap, M.: On the resilience of llm-based multi-agent collaboration with faulty agents. In: Proceedings of the 42nd International Conference on Machine Learning (2025)

2025
[35]

In: Proceedings of the 42nd International Conference on Machine Learning

Zhang, S., Yin, M., Zhang, J., Liu, J., Han, Z., Zhang, J., Li, B., Wang, C., Wang, H., Chen, Y., Wu, Q.: Which agent causes task failures and when? On automated failure attribution of LLM multi-agent systems. In: Proceedings of the 42nd International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 267, pp. 76583–76599 (2025)

2025
[36]

Zhang, G., Wang, J., Chen, J., Zhou, W., Wang, K., Yan, S.: Agentracer: Who is inducing failure in the llm agentic systems? arXiv preprint arXiv:2509.03312 (2025)

work page arXiv 2025
[37]

In: Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis, pp

Ma, X., Wang, Y., Wang, J., Xie, X., Wu, B., Li, S., Xu, F., Wang, Q.: Enhancing multi-agent system testing with diversity-guided exploration and adaptive critical state exploitation. In: Proceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis, pp. 1491–1503 (2024)

2024
[38]

In: International Conference on Learning Representations (ICLR) (2021)

Shridhar, M., Yuan, X., Côté, M.-A., Bisk, Y., Trischler, A., Hausknecht, M.: Alfworld: Aligning text and embodied environments for interactive learning. In: International Conference on Learning Representations (ICLR) (2021)

2021
[39]

Finance agent benchmark: Benchmarking llms on real-world financial research tasks.arXiv preprint arXiv:2508.00828,

Bigeard, A., Krishnan, R., Wu, S., Nashold, L.: Finance agent benchmark: Benchmarking llms on real-world financial research tasks. arXiv preprint arXiv:2508.00828 (2025)

work page arXiv 2025
[40]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

Shridhar, M., Thomason, J., Gordon, D., Bisk, Y., Han, W., Mottaghi, R., Zettlemoyer, L., Fox, D.: Alfred: A benchmark for interpreting grounded instructions for everyday tasks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10740–10749 (2020) 37

2020