Initial Steps in Integrating Large Reasoning and Action Models for Service Composition

Ilche Georgievski; Marco Aiello

REVIEW 3 major objections 5 minor 32 references

Initial Steps in Integrating Large Reasoning and Action Models for Service Composition

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that pairing Large Reasoning Models with Large Action Models in a three-layer architecture can automate the entire service-composition lifecycle from natural-language intent to executed, adapted workflows.

desk verdict A useful position paper with a solid survey and a plausible LRM-LAM architecture, but it assigns the reasoning-heavy adaptation tasks to LAMs despite its own stated LAM limitation. read the letter →

arxiv 2507.18775 v1 pith:ZIZMHULW submitted 2025-07-24 cs.AI cs.SE

Ilche Georgievski , Marco Aiello This is my paper

classification cs.AIcs.SE

keywords automatedservicecompositionLargeReasoningModelsActionlanguageservice-orientedcomputingworkflowadaptationarchitecturalframeworknatural-languageintent

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Automated service composition—finding, connecting, and running services to fulfil a user goal—has stalled at roughly Level 2 of a five-level automation scale: systems can discover and assemble services but cannot deeply understand requests, reason at ecosystem scale, or recover from execution failures. The paper's central claim is that the two newest language-model paradigms are complementary in exactly the right way: Large Reasoning Models (LRMs) can interpret natural-language intent and reason about service compatibility, while Large Action Models (LAMs) can call APIs, execute workflows, and adapt to failures. To exploit that, it proposes a three-layer architecture that assigns request analysis and discovery to LRMs and LAMs, composition to LRMs, and execution, monitoring, and adaptation to LAMs, with a coordination layer and a training phase that feed execution data back into the models. If the integration works, a user could state a high-level intention and the system would handle the full composition lifecycle without human intervention. The paper presents this as a conceptual blueprint and research roadmap rather than as an implemented or evaluated system.

What carries the argument

The central object is the LRM–LAM architectural framework, organised as three inference layers plus a coordination layer and a training phase. Layer 1 handles request specification and service discovery, using an LRM for request analysis and semantic understanding, a LAM for retrieving service metadata, and an LRM for service selection. Layer 2 is the composition layer, where an LRM plans candidate compositions, optimises and selects among them, and validates logical correctness. Layer 3 is the execution and adaptation layer, where a LAM turns the composition into a workflow specification, handles API interfacing, executes services, monitors them, and recovers from failures. The coordination layer maintains a shared representation of composition state between reasoning and execution and generates data for the training phase, turning the pipeline into a cyclic, self-improving system.

What would settle it

Build a prototype of the three-layer architecture using currently available public LRMs and LAMs, run it against a benchmark of heterogeneous service requests with injected API failures, and compare composition validity, success rate, and recovery time against existing LLM-based service-composition baselines; if the prototype's end-to-end success rate is no better than the baselines, or if Layer 3 cannot recover from injected failures, the proposed division of labour is not delivering the claimed automation.

Watch

Extended reading notes

Core claim

The discovery the paper wants to establish is that the bottleneck in automated service composition is not a missing component but a misaligned division of labour: current LLM-based approaches use one general-purpose model for both understanding and acting, and that model is shallow at reasoning and unreliable at grounded action. The paper's answer is an architectural separation in which the 'why'—interpreting the request, building a semantic picture of services, planning and validating the composition—is delegated to LRMs, and the 'how'—specifying, invoking, monitoring, and repairing execution—is delegated to LAMs. The three inference layers map directly onto the six phases of the service-composition pipeline, and the coordination layer closes the loop by routing execution results back into reasoning and by generating training data from both composition logic and failure patterns. The paper claims this integrated division of labour can cover all six phases, closing the execution-and-adaptation gap that the LLM-based state of the art leaves largely open.

Load-bearing premise

The framework assumes that current LRMs and LAMs can reliably perform the specific reasoning and execution tasks assigned to them in each layer; that premise is imported from general surveys of the model classes and is not tested on service-composition workloads.

Editorial extensions

If this is right

If the architecture works, service composition could move from Level 2 to Level 5 automation, with the system handling the entire lifecycle from natural-language request to executed and adapted composition.
The four listed challenges—context understanding, service integration, execution and adaptation, and reasoning at scale—would each be addressed by one side of the LRM–LAM split.
Execution failures and recovery strategies would become training data, so the system's composition and adaptation abilities would improve over successive cycles.
Existing LLM-based approaches for service discovery and composition could be repurposed as concrete starting points for realising individual layers.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the LRM–LAM separation proves effective, the same layered pattern could be applied to other end-to-end automation problems, such as business process orchestration or personal-assistant workflows, where semantic reasoning and API execution must interleave.
The paper leaves open how much of the coordination layer should be symbolic rather than model-driven; a likely testable extension is to insert a classical planner or formal verifier between Layer 2 and Layer 3 to catch invalid compositions before execution.
As model families converge toward combined reasoning-and-action capabilities, the paper's hard layer boundaries may blur; the architecture's value might then be less about which model class does what and more about the feedback loop that keeps execution data flowing into training.
A concrete benchmark extension would measure whether routing a single request through separate LRM and LAM calls is more reliable and cheaper than one general LLM agent doing the same job.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First thing to know: this is a genuine position paper, honest about having no prototype, and the conceptual architecture is plausible. But the internal division of labor has a real tension the authors don't address. They assign Service Failure Handling (Layer 3) to LAMs—root cause determination, failure type assessment, and selection among retry, replacement, reordering, and composition refinement. Yet Section 2.4 says LAMs 'often rely on learned patterns of behaviour and may have limited ability to reason and plan in complex situations,' and Figure 2 explicitly lists 'often weak at (high-level) reasoning' as a LAM limitation. Choosing between reordering and composition refinement is a planning task, not an execution task. So the architecture puts the reasoning-heavy part of adaptation on the model type it says is weak at reasoning. The coordination layer could in principle route this back to the LRM, but the text in Section 4.3 does not say that, and the claim of full-lifecycle coverage depends on this step.

What the paper does well: it lays out four challenges for automated service composition and maps existing LLM-based work onto the six-phase pipeline in Table 1. That table is a useful survey. The proposed three-layer framework with a coordination layer for feedback and training-data generation is a reasonable blueprint, and the coordination layer is a genuine addition over prior works. The authors also state clearly that feasibility and impact labels in Table 2 are approximate; treat them as editorial opinion.

The weaknesses beyond the LAM tension: no experiments, no formal verification, no prototype. That's acknowledged in Section 5, and for a position paper it's a limit rather than a fatal flaw. The definitions of LRM and LAM are imported from surveys, so the framework inherits whatever instability those taxonomies have. Self-citation is not a problem here—the cited prior works by the same group are relevant and legitimate.

Who benefits: researchers in service-oriented computing who want a starting point for integrating reasoning and action models. It is a roadmap, not a result. If a serious editor sends it to review, the authors should be asked to resolve the LAM reasoning assignment and either temper the 'full lifecycle' language or justify why LAMs can do failure diagnosis. I would accept it for peer review as a position paper, but not as anything stronger.

Referee Report

3 major / 5 minor

Summary. The paper argues that integrating Large Reasoning Models (LRMs) and Large Action Models (LAMs) can overcome the four challenges that currently limit automated service composition to around Level 2 automation. It provides background on LLMs, LRMs, and LAMs, reviews the state of the art in LLM-based service composition (Tables 1 and 2), and proposes a three-layer architectural framework (Request Analysis & Service Discovery, Service Composition, Service Execution & Adaptation) with a coordination layer and a training phase. The central claim is that this LRM-LAM synergy enables end-to-end automated service composition driven by natural language intent, covering the full lifecycle from request specification to execution and adaptation.

Significance. If the proposed division of labor is sound, this is a timely conceptual contribution that could serve as a roadmap for combining reasoning and action models in service-oriented computing. The paper correctly identifies Challenge 3 (execution and adaptation) as the least explored phase and makes a credible case for why a synergistic combination of LRMs and LAMs is a natural fit for the service composition pipeline. The architecture is clearly structured, and the explicit acknowledgement in Section 5 that no prototype exists and that evaluation is future work is honest. However, the significance is conditional on the internal consistency of the model-capability assignments and on the load-bearing assumption that current LRMs and LAMs can reliably perform the tasks assigned to them in a service composition context; the paper provides no evidence for this assumption beyond general survey statements.

major comments (3)

[§4.3 vs. §2.4 and Fig. 2] The assignment of Service Failure Handling to LAMs in Layer 3 contradicts the paper's own stated limitations of LAMs. The task includes determining root causes, assessing failure types, and generating recovery strategies such as reordering execution or composition refinement, all of which are reasoning and planning activities. Yet Section 2.4 states that LAMs 'often rely on learned patterns of behaviour and may have limited ability to reason and plan in complex situations,' and Figure 2 explicitly lists 'Often weak at (high-level) reasoning' as a LAM limitation. Since Challenge 3 is identified in Section 3 as the least explored phase, and the paper provides no evidence that current LAMs can perform this diagnostic and recovery-selection reasoning in service environments, the framework's claim to cover the full service composition lifecycle is not supported by its own assumptions. Please either reassign the diagnosis and strategy-generation portions of failure handling to Layer 2 (LRM) or to a collaborative LRM-LAM loop, or provide an argument for why LAMs can handle these reasoning tasks despite the stated limitation.
[§4 (Coordination layer) and §4.3] The control flow for failure handling and adaptation is ambiguous. The coordination layer is described as routing 'execution results from Layer 3 back to Layer 2, enabling iterative and reflective refinement based on real-world performance,' which suggests that composition refinement is an LRM responsibility. However, Section 4.3 assigns Service Failure Handling to Layer 3 (LAM) and includes 'composition refinement' among the LAM-generated recovery strategies. The reader cannot determine whether the LAM alone decides to reorder or refine the composition, whether it must consult the LRM, or whether the coordination layer mediates. This ambiguity matters because the paper's core contribution is precisely the division of labor between LRMs and LAMs. Please specify the data and control flow for failure detection, diagnosis, strategy selection, and refinement, including which component has authority in each step.
[§5 (Conclusion and Outlook)] The paper's central claim of achieving 'full lifecycle coverage' and potentially transforming service composition into a 'fully automated, user-friendly process' is not backed by empirical evidence, a formal specification, or a prototype. The authors acknowledge this in Section 5, where they list 'development of prototypes' and 'evaluation and adaptation of existing LRMs and LAMs' as future work. For a position paper this is acceptable, but the strength of the language in the abstract and introduction goes well beyond what is demonstrated. I recommend tempering the claims to match the conceptual nature of the contribution, or adding a more detailed feasibility analysis that maps each layer's tasks to known capabilities and limitations of specific existing LRMs and LAMs.

minor comments (5)

[§3, Table 2] The feasibility and impact labels in Table 2 are described as 'approximate and based on our interpretation' without a transparent rubric. Since the table is used to characterize the state of the art, adding a short definition for each label value (e.g., what qualifies as 'High' vs 'Medium' feasibility) would improve reproducibility and reduce subjectivity.
[Abstract] The abstract contains a typo: 'buildin g' should be 'building'.
[References] Reference [21] contains a duplicated 'arXivarXiv' in the journal field. Please fix the citation formatting.
[§2.1 and Fig. 1] The paper mentions the automation levels (Level 0 to Level 5) in the Introduction but never defines the intermediate levels explicitly. A one-sentence characterization of each level would help the reader place the challenges and the proposed framework in the context of the automation continuum.
[Fig. 3] The legend of Figure 3 is empty in the provided manuscript; the figure includes a 'Legend' box but no legend entries. Please ensure the legend items (e.g., 'Automated Task', 'Data flow', 'Layer boundary') are visible in the final version.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the LRM-LAM architecture is a proposal, not a fitted or self-referential derivation.

full rationale

The paper makes no quantitative prediction and fits no parameters; its central claim is the proposed three-layer LRM/LAM architecture for automated service composition. The argument proceeds by mapping pipeline phases to model capabilities described in Sections 2.3 and 2.4 (e.g., LRMs for UC Request Analysis and Composition Planning; LAMs for Service Interfacing and Execution). This mapping is not derived from an equation, a fitted value, or a uniqueness theorem; it is a design proposal. Self-citations [2], [3], [14], and [15] are used to characterize the state of the art and to position Challenge 3 as unexplored, but those sources are not the load-bearing justification for the architecture, and Section 5 explicitly admits that existing works do not implement the proposed framework. The LAM weakness noted in Section 2.4 ('often rely on learned patterns of behaviour and may have limited ability to reason and plan in complex situations') versus the Service Failure Handling task assigned to LAMs in Section 4.3 is an internal consistency and feasibility concern, not a circularity: it does not make the architecture's output equivalent to its input. Therefore the circularity score is 0.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper introduces no fitted parameters and no new entities in the sense of physics or mathematics. Its load-bearing assumptions are domain assumptions about the capabilities of LRMs and LAMs and the completeness of the pipeline model, plus an ad hoc thesis that the division of labor meets the four challenges. All remain unvalidated.

assumptions (4)

domain assumption LRMs provide deep semantic reasoning and planning capabilities.
Taken from surveys [4, 26] and asserted as a basis for Layer 1 and Layer 2 tasks, with no validation in the service composition context.
domain assumption LAMs can reliably execute and adapt actions in dynamic environments.
Taken from [22] and used to justify Layer 3 tasks, but no evidence is given for service composition workloads.
domain assumption The six-phase pipeline is a comprehensive model of automated service composition.
Adapted from [13] and used as the backbone of the framework.
ad hoc to paper The four challenges are the critical barriers to Level 5 automation and the LRM-LAM division addresses them.
This complementarity argument is the paper's main thesis and is not empirically supported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Initial Steps in Integrating Large Reasoning and Action Models for Service Composition." pith.science (2026). https://pith.science/paper/ZIZMHULW

@misc{pith2026250718775,
  author       = {Pith},
  title        = {Pith review of: Initial Steps in Integrating Large Reasoning and Action Models for Service Composition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZIZMHULW}},
  note         = {Machine review of arXiv:2507.18775}
}

read the original abstract

Service composition remains a central challenge in building adaptive and intelligent software systems, often constrained by limited reasoning capabilities or brittle execution mechanisms. This paper explores the integration of two emerging paradigms enabled by large language models: Large Reasoning Models (LRMs) and Large Action Models (LAMs). We argue that LRMs address the challenges of semantic reasoning and ecosystem complexity while LAMs excel in dynamic action execution and system interoperability. However, each paradigm has complementary limitations - LRMs lack grounded action capabilities, and LAMs often struggle with deep reasoning. We propose an integrated LRM-LAM architectural framework as a promising direction for advancing automated service composition. Such a system can reason about service requirements and constraints while dynamically executing workflows, thus bridging the gap between intention and execution. This integration has the potential to transform service composition into a fully automated, user-friendly process driven by high-level natural language intent.

Figures

Figures reproduced from arXiv: 2507.18775 by the authors.

**Figure 1.** Automated service composition pipeline with illustrative example from the personal assistance domain. request is specified. This can be a direct input from the user, or the system can observe and specify the user’s objectives. In the second phase, the system searches and identifies available services that could potentially fulfill parts of the request. Traditionally, this phase involves querying service registries, … view at source ↗

**Figure 2.** Language Model Types LLMs to excel in understanding human language and generating coherent responses, such as text, code, and other similar outputs. The question of whether LLMs have the potential to solve complex real-world problems through abstract and logical reasoning has been of particular interest and debate. A notable concept in this context is the concept of thought, which is a sequence of tokens that repre… view at source ↗

**Figure 3.** Architectural Framework [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 23 canonical work pages

[1]

In: International Conference on Service-Oriented Computing

Aiello, M.: A challenge for the next 50 years of automated s ervice composition. In: International Conference on Service-Oriented Computing. pp. 635–643. Springer (2022)

work page 2022
[2]

IEEE Transactions on Services Computing (2025)

Aiello, M.: A paradigm shift in service research: The case of service composition. IEEE Transactions on Services Computing (2025)

work page 2025
[3]

Service Ori- ented Computing and Applications 17(4), 233–238 (2023)

Aiello, M., Georgievski, I.: Service composition in the C hatGPT era. Service Ori- ented Computing and Applications 17(4), 233–238 (2023)

work page 2023
[4]

Besta, M., Barth, J., Schreiber, E., Kubicek, A., Catarin o, A., Gerstenberger, R., Nyczyk, P., Iﬀ, P., Li, Y., Houliston, S., et al.: Reasoning l anguage models: A blueprint. Tech. Rep. 2501.11223, arXiv (2025)

arXiv 2025
[5]

In: IEEE International Conf erence on Web Services

Bianchini, D., Garda, M., Melchiori, M., Rula, A.: Levera ging large language mod- els for data service discovery. In: IEEE International Conf erence on Web Services. pp. 1097–1099. IEEE (2024)

work page 2024
[6]

Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X ., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., et al.: Rt-2: Visio n-language-action mod- els transfer web knowledge to robotic control. Tech. Rep. 23 07.15818, arXiv (2023)

work page 2023
[7]

Chen, B., Shu, C., Shareghi, E., Collier, N., Narasimhan, K., Yao, S.: Fireact: Toward language agent ﬁne-tuning. Tech. Rep. 2310.05915, a rXiv (2023) Title Suppressed Due to Excessive Length 15

arXiv 2023
[8]

Dong, Q., Li, L., Dai, D., Zheng, C., Ma, J., Li, R., Xia, H., Xu, J., Wu, Z., Liu, T., et al.: A survey on in-context learning. Tech. Rep. 2303. 18223, arXiv (2022)

work page 2022

Show all 32 references

[9]

In: IEEE/ACM International Conference on Automated Softwa re Engineering

Huang, Q., Wan, Z., Xing, Z., Wang, C., Chen, J., Xu, X., Lu, Q.: Let’s chat to ﬁnd the APIs: Connecting human, LLM and knowledge graph thro ugh AI chain. In: IEEE/ACM International Conference on Automated Softwa re Engineering. pp. 471–483. IEEE (2023)

2023
[10]

Kambhampati, S., Valmeekam, K., Guan, L., Verma, M., Ste chly, K., Bhambri, S., Saldyt, L., Murthy, A.: LLMs can’t plan, but can help plannin g in LLM-modulo frameworks. Tech. Rep. 2402.01817, arXiv (2024)

2024 arXiv
[11]

Cluste r Computing 27(4), 4035–4061 (2024)

Kotstein, S., Decker, C.: Restberta: A transformer-bas ed question answering ap- proach for semantic search in Web API documentation. Cluste r Computing 27(4), 4035–4061 (2024)

2024
[12]

, Chang, Y., Wang, Q.: Screenagent: A vision language model-driven computer cont rol agent

Niu, R., Li, J., Wang, S., Fu, Y., Hu, X., Leng, X., Kong, H. , Chang, Y., Wang, Q.: Screenagent: A vision language model-driven computer cont rol agent. Tech. Rep. 2402.07945, arXiv (2024)

2024 arXiv
[13]

In: Proceedings of the Fourth International Confere nce on Web Information Systems Engineering, 2003

Papazoglou, M.P.: Service-oriented computing: Concep ts, characteristics and direc- tions. In: Proceedings of the Fourth International Confere nce on Web Information Systems Engineering, 2003. WISE 2003. pp. 3–12. IEEE (2003)

2003
[14]

In: Intern ational Conference on Service-Oriented Computing

Pesl, R.D., Mombrey, C., Klein, K., Zyberaj, D., Georgie vski, I., Becker, S., Herzwurm, G., Aiello, M.: Compositio Prompto: An architect ure to employ large language models in automated service computing. In: Intern ational Conference on Service-Oriented Computing. pp. 276–28...

2024
[15]

In: Internati onal Conference on Service- Oriented Computing

Pesl, R.D., Stötzner, M., Georgievski, I., Aiello, M.: U ncovering LLMs for service- composition: challenges and opportunities. In: Internati onal Conference on Service- Oriented Computing. pp. 39–48. Springer (2023)

2023
[16]

Advances in Neural Information Pro cessing Systems 36, 68539–68551 (2023)

Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lom eli, M., Hambro, E., Zettle- moyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Pro cessing Systems 36, 68539–68551 (2023)

2023
[17]

Nature 623(7987), 493–498 (2023)

Shanahan, M., McDonell, K., Reynolds, L.: Role play with large language models. Nature 623(7987), 493–498 (2023)

2023
[18]

Song, Y., Xiong, W., Zhu, D., Wu, W., Qian, H., Song, M., Hu ang, H., Li, C., Wang, K., Yao, R., et al.: Restgpt: Connecting large languag e models with real- world restful APIs. Tech. Rep. 2306.06624, arXiv (2023)

2023 arXiv
[19]

In: NeurIPS 2022 Foundation Models for Decision Ma king Workshop (2022)

Valmeekam, K., Olmo, A., Sreedharan, S., Kambhampati, S .: Large language mod- els still can’t plan (a benchmark for LLMs on planning and rea soning about change). In: NeurIPS 2022 Foundation Models for Decision Ma king Workshop (2022)

2022
[20]

Trans- actions on Machine Learning Research (2025)

Valmeekam, K., Stechly, K., Gundawar, A., Kambhampati, S.: A systematic eval- uation of the planning and scheduling abilities of the reaso ning model o1. Trans- actions on Machine Learning Research (2025)

2025
[21]

Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu , Y., Fan, L., Anand- kumar, A.: Voyager: An open-ended embodied agent with large language models. Tech. Rep. 2305.16291, arXivarXiv (2023)

2023 arXiv
[22]

Wang, L., Yang, F., Zhang, C., Lu, J., Qian, J., He, S., Zha o, P., Qiao, B., Huang, R., Qin, S., et al.: Large action models: From inception to im plementation. Tech. Rep. 2412.10047, arXiv (2024)

2024 arXiv
[23]

In: Inter national Conference on Service-Oriented Computing

Wang, X., Zhou, P., Wang, Y., Liu, X., Liu, J., Wu, H.: Serv icebert: A pre-trained model for web service tagging and recommendation. In: Inter national Conference on Service-Oriented Computing. pp. 464–478 (2021) 16 Georgievski and Aiello

2021
[24]

Wei, J., Tay, Y., Bommasani, R., Raﬀel, C., Zoph, B., Borg eaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., et al.: Emergent abilit ies of large language models. Tech. Rep. 2206.07682, arXiv (2022)

2022 arXiv
[25]

Advances in neural information processing systems 35, 24824–24837 (2022)

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi , E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning i n large language models. Advances in neural information processing systems 35, 24824–24837 (2022)

2022
[26]

Xu, F., Hao, Q., Zong, Z., Wang, J., Zhang, Y., Wang, J., La n, X., Gong, J., Ouyang, T., Meng, F., et al.: Towards large reasoning models : A survey of rein- forced reasoning with large language models. Tech. Rep. 250 1.09686, arXiv (2025)

2025
[27]

Advances in Neural Information Processing Systems 35, 20744–20757 (2022)

Yao, S., Chen, H., Yang, J., Narasimhan, K.: Webshop: Tow ards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems 35, 20744–20757 (2022)

2022
[28]

Advances in neural information processing systems 36, 11809–11822 (2023)

Yao, S., Yu, D., Zhao, J., Shafran, I., Griﬃths, T., Cao, Y ., Narasimhan, K.: Tree of thoughts: Deliberate problem solving with large languag e models. Advances in neural information processing systems 36, 11809–11822 (2023)

2023
[29]

In: In ternational Conference on Learning Representations (2023)

Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan , K., Cao, Y.: React: Synergizing reasoning and acting in language models. In: In ternational Conference on Learning Representations (2023)

2023
[30]

Zeng, A., Liu, M., Lu, R., Wang, B., Liu, X., Dong, Y., Tang , J.: Agenttuning: Enabling generalized agent abilities for LLMs. Tech. Rep. 2 310.12823, arXiv (2023)

2023
[31]

In: 2024 5th International Symposium on Comput er Engineering and Intelligent Communications (ISCEIC)

Zhang, L., Zhao, X., Wang, L.: Automatic service composi tion agent based on large model. In: 2024 5th International Symposium on Comput er Engineering and Intelligent Communications (ISCEIC). pp. 310–316. IEEE (2 024)

2024
[32]

Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., M in, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language mod els. Tech. Rep. 2303.18223, arXiv (2023)

2023 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

In: International Conference on Service-Oriented Computing

Aiello, M.: A challenge for the next 50 years of automated s ervice composition. In: International Conference on Service-Oriented Computing. pp. 635–643. Springer (2022)

work page 2022

[2] [2]

IEEE Transactions on Services Computing (2025)

Aiello, M.: A paradigm shift in service research: The case of service composition. IEEE Transactions on Services Computing (2025)

work page 2025

[3] [3]

Service Ori- ented Computing and Applications 17(4), 233–238 (2023)

Aiello, M., Georgievski, I.: Service composition in the C hatGPT era. Service Ori- ented Computing and Applications 17(4), 233–238 (2023)

work page 2023

[4] [4]

Besta, M., Barth, J., Schreiber, E., Kubicek, A., Catarin o, A., Gerstenberger, R., Nyczyk, P., Iﬀ, P., Li, Y., Houliston, S., et al.: Reasoning l anguage models: A blueprint. Tech. Rep. 2501.11223, arXiv (2025)

arXiv 2025

[5] [5]

In: IEEE International Conf erence on Web Services

Bianchini, D., Garda, M., Melchiori, M., Rula, A.: Levera ging large language mod- els for data service discovery. In: IEEE International Conf erence on Web Services. pp. 1097–1099. IEEE (2024)

work page 2024

[6] [6]

Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X ., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., et al.: Rt-2: Visio n-language-action mod- els transfer web knowledge to robotic control. Tech. Rep. 23 07.15818, arXiv (2023)

work page 2023

[7] [7]

Chen, B., Shu, C., Shareghi, E., Collier, N., Narasimhan, K., Yao, S.: Fireact: Toward language agent ﬁne-tuning. Tech. Rep. 2310.05915, a rXiv (2023) Title Suppressed Due to Excessive Length 15

arXiv 2023

[8] [8]

Dong, Q., Li, L., Dai, D., Zheng, C., Ma, J., Li, R., Xia, H., Xu, J., Wu, Z., Liu, T., et al.: A survey on in-context learning. Tech. Rep. 2303. 18223, arXiv (2022)

work page 2022