Pith. sign in

REVIEW 3 major objections 6 minor 41 references

Z-COPA turns manual 0D network planning into multi-agent graph topology–parameter co-optimization and reports the best forward and inverse air-system scores among its agent baselines.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 07:47 UTC pith:LZZJ3W4C

load-bearing objection Solid multi-agent + solver pipeline for 0D network co-planning; real SAS CAD cases and honest ablations, but the big quality claims are only vs agent baselines and the 0D fidelity ceiling is real. the 3 major comments →

arxiv 2607.10994 v1 pith:LZZJ3W4C submitted 2026-07-13 cs.LG

A Multi-Agent Framework for Zero-Dimensional Reduced-Order Model Planning

Bingteng Sun , Hao Yin , Yiling Chen , Renjie Xiao , Lei Xie , Shanyou Wang , Ruonan Wang , Shubao Chen
show 4 more authors
Qingzong Xu Lin Lu Qiang Du Junqiang Zhu
This is my paper
classification cs.LG
keywords Large Language Model AgentZero-Dimensional Reduced-Order Model PlanningRetrieval-Augmented GenerationTopology-Parameter CO-OptimizationAero-Engine Air Systems DesignMulti-Agent FrameworkGraph Structure OptimizationMILP-Guided Navigation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Zero-dimensional reduced-order models sit at the start of many complex equipment design chains, yet their network topology and parameters are still largely chosen by experts or by local parameter tuners. This paper claims that a multi-agent system can replace that empirical loop: it encodes 0D flow networks as attributed graphs, grounds agents with ontology-driven retrieval (SAGE), and searches mixed topology and parameters with a MILP-guided navigator (MGN) under a physics solver. On two real aero-engine secondary-air systems plus IEEE power-grid and water-network benchmarks, the framework reports superior task completion, the best forward and reverse air-system scores among the paper’s agent baselines, a 55.46% active-power-loss cut on IEEE 69-bus, and a 90.45% cost reduction on the Two-Loop water network. A sympathetic reader would care because the claim is that broader topological space becomes searchable and that design intent can move from drawings and targets to solver-validated architectures with far less hand-crafted setup.

Core claim

The paper establishes that 0D ROM planning can be cast as a constrained topology–parameter co-optimization problem on an attributed design graph, and that a multi-agent architecture (Z-COPA) with SAGE knowledge grounding and MGN hybrid search can solve both forward parameter refinement on fixed topologies and inverse topology–parameter generation from incomplete engineering inputs, outperforming single-agent ReAct/CoT and multi-agent NSGA-II baselines on the reported air-system, power, and water cases.

What carries the argument

The dedicated graph representation of 0D flow-network topology, paired with the Symbolic Action Graph Engine (SAGE) for ontology-driven multimodal retrieval and the MILP-Guided Navigation (MGN) optimizer that linearizes local regions, screens discrete topology with continuous parameters, then feeds elite candidates into Pareto evolution under a graph-based physics solver (Z-GPSolver).

Load-bearing premise

The method assumes the underlying zero-dimensional physics models and solvers are accurate enough that solver-backed objective gains match real engineering performance.

What would settle it

Take a Z-COPA inverse-design topology–parameter set for a real secondary-air system and re-evaluate seal leakage, total bleed flow, and cavity pressure nonuniformity under instrumented hardware or high-fidelity 3D CFD; if the reported multi-objective gains reverse or disappear, the central claim does not hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • 0D air-system architectures can be explored over wider topological spaces without relying only on expert trial-and-error.
  • Forward and inverse secondary-air-system design can reach higher multi-objective scores under the paper’s agent baselines than single-agent or NSGA-II multi-agent setups.
  • The same agent–solver workflow transfers to power-distribution reconfiguration and water-network pipe sizing with near-reference losses and costs.
  • Iteration cycles shorten because perception, retrieval, candidate generation, and solver validation form a closed, auditable loop.
  • Formulating objectives and constraints needs less manual intervention once the structured graph state and ontology bindings are in place.

Where Pith is reading between the lines

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

  • If the graph contract and multi-agent loop hold up, the same CO-planning pattern could apply to other 0D networks such as HVAC, process flowsheets, or gas pipelines without rewriting the agent stack for each domain.
  • Feeding 3D field corrections into the 0D loop, as the authors sketch for future work, would make the method a natural bridge from early topology search to later high-fidelity refinement.
  • The results suggest engineering multi-agent systems win more from strict tool contracts, solver authority, and structured state than from unconstrained LLM reasoning alone.
  • Public benchmarks that mix topology edits with hard radiality, connectivity, and pressure constraints may become standard stress tests for agent design systems beyond language-only tasks.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript proposes Z-COPA, a multi-agent LLM framework for zero-dimensional reduced-order model (0D ROM) planning that couples a Symbolic Action Graph Engine (SAGE) for ontology-driven multimodal RAG with a MILP-Guided Navigation (MGN) optimizer and a graph PINN-style solver (Z-GPSolver). It formulates 0D planning as constrained topology–parameter co-optimization x=(T,P), supporting forward design (fixed topology, parameter search) and inverse design (joint topology–parameter search) under solver-backed physical constraints. Validation covers two real aero-engine secondary-air-system (SAS) CAD-driven cases, IEEE 33/69 distribution reconfiguration, and Two-Loop / New York Tunnels water-network design. Ablations isolate SAGE, ReAct vs Compact CoT, multi-agent decomposition, and MGN vs NSGA-II. Reported outcomes include best agent-baseline scores on SAS forward/inverse tasks, a 55.46% active-power-loss reduction on IEEE 69-bus, and large cost reductions on water benchmarks, with Z-GPSolver accelerating pressure-correction solves while keeping mean relative pressure error ~0.55%.

Significance. If the claims hold under stronger numerical controls, the work is a substantial systems contribution at the intersection of LLM multi-agent orchestration and engineering network design. Strengths include: (i) a concrete graph intermediate representation that makes CAD-to-solver planning inspectable; (ii) solver-backed closed-loop evaluation rather than text-only design; (iii) ablations that separate retrieval, reasoning format, multi-agent structure, and optimizer family; (iv) transfer across SAS, power, and water domains with external simulators (AC power flow, EPANET/WNTR); and (v) explicit acknowledgment that final quality is bounded by 0D model fidelity. The paper does not claim a new physical theory; its value is an automated, auditable co-planning stack for long-horizon 0D topology–parameter design.

major comments (3)
  1. §2.1 Baselines and §2.2.3 / Abstract quantitative claims: The central claim that Z-COPA “disrupts” 0D planning and delivers superior / near-reference quality is only partially isolated. All five baselines are agent ablations (Single ReAct/CoT ± SAGE + NSGA-II; Multi-Agent + SAGE + NSGA-II; Multi-Agent + SAGE + MGN). Classical topology optimizers are excluded once a solver-ready graph exists, yet IEEE 33/69 reconfiguration and Two-Loop/NYT pipe sizing have established GA/MILP/heuristic reference optima. The reported 55.46% IEEE 69 loss cut and 90.45% Two-Loop cost cut are relative to the paper’s initial designs / agent baselines (and small gaps to “reference baselines” in Fig. 5c), not under matched evaluation budgets against pure numerical topology search on the same fixed graph and solver. Without that control, MGN’s contribution to solution quality versus multi-agent orchestration as a
  2. §4.1.3 MGN Phases 1–2 and the local-linearity axiom: MGN rests on partitioning the design space into sub-regions where objectives and physical responses are treated as approximately linear so that first-order expansions feed a MILP over continuous x and discrete topology indicators z. For SAS inverse design and power/water topology edits, the true response is nonlinear (AC power flow, hydraulic head–loss, seal/cavity flow). The manuscript does not report residual linearization error, sub-region validity diagnostics, or failure cases when the linear MILP screen rejects or promotes candidates that the full solver later contradicts. Because MGN is the load-bearing optimizer distinguishing B5 from Multi-Agent+NSGA-II, this approximation needs quantitative stress-testing on at least one nonlinear benchmark with matched candidate budgets.
  3. §2.1 Evaluation metrics / Appendix A.5 Success Rate vs solution quality: SR is defined permissively as completing the agent–tool workflow and returning a solver-validated design with extractable objectives, explicitly to assess orchestration reliability rather than design quality. Figures 3b and 4b and the abstract’s “best performance” language mix SR, objective-wise improvements, and a reporting-only Total Score (SAS weights 0.5/0.3/0.2). The paper should report, for each case–baseline, the number of independent runs, SR, mean±std of each physical objective, and constraint-violation rates separately, and avoid letting the scalar Total Score or SR stand in for Pareto quality when claiming state-of-the-art design performance.
minor comments (6)
  1. Abstract and Conclusions: “globally optimal air system architectures” overstates what multi-objective constrained search with local linearization and NSGA-II can guarantee; prefer “high-quality / Pareto-improving under solver constraints.”
  2. Fig. 3–5: Several panels compress workflow traces, quantitative bars, and design evidence; enlarge fonts and separate agent traces from objective tables for readability.
  3. Notation: The optimization statement uses L(S(T,P), R_target) and g(T,P)≤0 early, but later SAS objectives are named seal leakage / total flow / pressure_cv without a single consistent symbol table linking them to L.
  4. Appendix B.1–B.2: MGN quotas and Z-GPSolver hyperparameters are listed; a short sensitivity note (or fixed seeds already used) would help reproducibility claims.
  5. Typos / polish: “W Writing-review” in CRediT; occasional duplicated phrasing in Introduction on multi-agent frameworks; ensure arXiv ID / citation consistency for concurrent RAG/agent works.
  6. Hardware note (§2.1): Remote kimi-k2.6 API dependence should be stated as a reproducibility constraint (model version, tool-call interface) alongside the local workstation specs.

Circularity Check

0 steps flagged

No significant circularity: Z-COPA is an agent-orchestrated engineering optimizer whose objectives are scored by external physics solvers, not by quantities fitted to the reported targets.

full rationale

This is a systems/methods paper, not a first-principles derivation. The load-bearing optimization statement (x* = arg min L(S(T,P), R_target) s.t. g ≤ 0) is a standard constrained co-optimization formulation; S is an external 0D/AC-power-flow/EPANET solver (or Z-GPSolver trained to residual consistency L_res = ||J Δp̂ + r||²), not a quantity defined from the reported improvements. MGN’s three phases (local linearization from Cartesian samples → MILP screening → NSGA-II elite injection) are algorithmic constructions that generate candidates; they do not redefine the physical objectives. The paper explicitly states that the scalar Total Score (weights 0.5/0.3/0.2 on SAS metrics) is used only for reporting and cross-baseline comparison, not as the internal multi-objective search objective. Benchmarks (IEEE 33/69, Two-Loop, New York Tunnels) and reference gaps are taken from the external literature (Baran–Wu, Alperovits–Shamir, Savic–Walters), not from self-cited uniqueness theorems. Success requires solver-backed tool evidence rather than text-only claims. No self-definitional loop, fitted-input-as-prediction, load-bearing self-citation uniqueness claim, or renaming of a known closed-form result appears in the derivation chain. Experimental-design concerns (agent ablations only; classical topology optimizers excluded once the graph is fixed) affect comparative strength, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 4 invented entities

The central claim rests less on new physics than on engineering assumptions: that 0D network models plus solver feedback are adequate design oracles; that ontology-bound retrieval sufficiently grounds LLM agents; and that local linear MILP approximations plus evolutionary search can navigate mixed topology–parameter spaces. Many free hyperparameters control MGN/Z-GPSolver behavior. Invented entities are methodological modules, not new physical objects.

free parameters (4)
  • SAS TotalScore reporting weights (w_SLS, w_TFS, w_PBS)
    Fixed at 0.5/0.3/0.2 for cross-baseline scalar ranking; not the internal optimizer objective, but they shape the headline ‘best performance’ comparisons.
  • MGN perception/action/integration quotas and schedules
    Table of hyperparameters (local sample pool 12, expansion 12, landscape passes 8, guided recombination 0.90, frozen-variable ratio 0.80, gateway budgets 20/5/15, seed 42, etc.) chosen by authors and load-bearing for reported search quality.
  • Z-GPSolver training and correction hyperparameters
    Hidden dim 64, 3 message passes, relaxation scale 0.005, init scale 1e-4, batch 16, lr 1.5e-4, weight decay 1e-4; fitted/selected for the air-system acceleration claims.
  • Critical boundary-flow floor
    Constraint threshold set to 0.95 of baseline in MGN selection; directly affects feasibility filtering of candidates.
axioms (4)
  • domain assumption 0D node–link models preserve the core mass/momentum/energy transfer needed for early architecture decisions.
    Stated throughout Introduction and Methods; Conclusions admit oversimplified component models can make optimized designs deviate from reality.
  • ad hoc to paper Within each local design sub-region, objectives and physical responses may be treated as approximately linear for MILP construction.
    MGN Phase 1 explicitly assumes local linearity after Cartesian sampling to build first-order expansions for MILP.
  • domain assumption Ontology-guided multimodal retrieval plus role-decomposed agents can expose legal variables/actions without the LLM implementing the physics solver.
    Core architectural premise of SAGE and the multi-agent harness in Sections 4.1–4.2 and cross-domain case discussion.
  • domain assumption Solver-backed residual/feasibility checks are sufficient final arbiters of design validity for reported success and quality metrics.
    Evaluation protocol counts success only with tool-supported generation, selection, and solver validation.
invented entities (4)
  • Z-COPA multi-agent co-planning architecture no independent evidence
    purpose: Orchestrate representation, design, and reflection agents for forward/inverse 0D ROM planning.
    Paper-specific system; evidence is internal experiments, not independent external deployments.
  • SAGE (Symbolic Action Graph Engine) no independent evidence
    purpose: Ontology-driven multimodal RAG over documents, solver code, and cases to ground agent decisions.
    New module name/architecture in this paper; no external independent validation cited.
  • MGN (MILP-Guided Navigation) no independent evidence
    purpose: Hybrid discrete–continuous candidate generation via local linearization, MILP screening, and Pareto evolution.
    Author-defined optimizer; performance claims come only from this paper’s ablations.
  • Z-GPSolver no independent evidence
    purpose: Graph-learned SPD pressure-correction assistant inside the 0D nonlinear solve loop.
    Presented as a closed-loop learned solver component with internal timing/accuracy tests only.

pith-pipeline@v1.1.0-grok45 · 30371 in / 3641 out tokens · 36710 ms · 2026-07-14T07:47:31.259070+00:00 · methodology

0 comments
read the original abstract

Zero-dimensional reduced-order models (0D ROMs) are central to multi-dimensional design workflows for high-end complex equipment. However, the planning process currently relies on manual expertise, limiting topological exploration and prolonging iterations. Even traditional optimization methods such as Genetic Algorithms (GA) are typically confined to local parameter tuning. Although Large Language Model (LLM) agents have shown promise in exploring large sample spaces, and frameworks such as Chain of Thought (CoT) and Reason and Act (ReAct) improve reasoning reliability, while Retrieval-Augmented Generation (RAG) overcomes domain knowledge barriers, a single agent still falls short for the long-horizon and highly coupled nature of complex 0D ROM planning. This paper proposes the Zero-dimensional reduced-order model CO-Planning framework (Z-COPA), a multi-agent architecture featuring a Symbolic Action Graph Engine (SAGE) and a MILP-Guided Navigation (MGN) optimizer. Its core innovation is a dedicated graph representation method that accurately encodes the 0D flow network topology, converting the empirical planning process into a rigorous graph structure optimization problem. We validate the forward and inverse design capabilities and generalization performance of Z-COPA on two real aircraft engine secondary-air systems, two IEEE power-distribution reconfiguration benchmarks, and two water-distribution network benchmarks. The results show superior task completion quality, obtaining the best performance in both forward and reverse design of air systems. Z-COPA disrupts the traditional 0D model planning paradigm, providing a new technical approach for exploring broader topological space and achieving highly automated, globally optimal air system architectures.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 14 canonical work pages

  1. [1]

    Applied Energy365, 123234 (2024) https://doi.org/10.1016/j.apenergy.2024

    Culubret, S., Rubio, M.A., Sanchez, D.G., Urquia, A.: Performance uniformity analysis in polymer electrolyte fuel cell using long-term dynamic simulation. Applied Energy365, 123234 (2024) https://doi.org/10.1016/j.apenergy.2024. 123234 28

    doi:10.1016/j.apenergy.2024 2024

  2. [2]

    Physics of Fluids38(3) (2026) https://doi

    Roknujjaman, M., Rahman, M.M., Rahman, M.M., Hasan, M.M., Baki, A.B.M., Zhao, M.: Pulsatile flow dynamics in biomedical engineering: A comprehensive review of methods and applications. Physics of Fluids38(3) (2026) https://doi. org/10.1063/5.0318753

    work page doi:10.1063/5.0318753 2026

  3. [3]

    Duronio, F., De Vita, A., Allocca, L., Montanaro, A., Ranieri, S., Villante, C.: Cfd numerical reconstruction of the flash boiling gasoline spray morphology1 (2020) https://doi.org/10.4271/2020-24-0010

    work page doi:10.4271/2020-24-0010 2020

  4. [5]

    International Journal of Turbomachinery, Propulsion and Power10(4), 37 (2025) https://doi.org/10.3390/ijtpp10040037

    Pinardi, A., Mushtaq, N., Gaetani, P.: Development of a reduced order model for turbine blade cooling design. International Journal of Turbomachinery, Propulsion and Power10(4), 37 (2025) https://doi.org/10.3390/ijtpp10040037

    work page doi:10.3390/ijtpp10040037 2025

  5. [6]

    International Jour- nal of Hydrogen Energy48(65), 25469–25486 (2023) https://doi.org/10.1016/j

    Wang, C., Zhou, D., Xiao, W., Shui, C., Ma, T., Chen, P., Hao, J., Yan, J.: Research on the dynamic characteristics of natural gas pipeline network with hydrogen injection considering line-pack influence. International Jour- nal of Hydrogen Energy48(65), 25469–25486 (2023) https://doi.org/10.1016/j. ijhydene.2023.03.298

    doi:10.1016/j 2023

  6. [7]

    Aerospace (2026)

    Yan, J., Chen, Z., Huang, J., Zhou, W.: A non-iterative calculation method for zero-dimensional nozzle model of gas turbine engine. Aerospace (2026)

    2026

  7. [8]

    Journal of the Global Power and Propulsion Society8, 310–322 (2024) https://doi.org/ 10.33737/jgpps/191167

    Zou, W., Song, Z., Wang, B., Wen, M., Zheng, X.: An efficient multi-fidelity simulation approach for performance prediction of adaptive cycle engines. Journal of the Global Power and Propulsion Society8, 310–322 (2024) https://doi.org/ 10.33737/jgpps/191167

    work page doi:10.33737/jgpps/191167 2024

  8. [9]

    Technical University of Denmark (DTU) Publications (2027)

    Wu, J., Lind, M.: Combining empirical knowledge and first principles qualitative and quantitative models in engineering design. Technical University of Denmark (DTU) Publications (2027)

    2027

  9. [10]

    Journal of Water Resources Planning and Management146(12) (2020) https://doi.org/ 10.1061/(asce)wr.1943-5452.0001299

    Liu, H., Shoemaker, C.A., Jiang, Y., Fu, G., Zhang, C.: Preconditioning water distribution network optimization with head loss–based design method. Journal of Water Resources Planning and Management146(12) (2020) https://doi.org/ 10.1061/(asce)wr.1943-5452.0001299

    work page doi:10.1061/(asce)wr.1943-5452.0001299 2020

  10. [11]

    Water14(6), 851 (2022) https://doi.org/10.3390/w14060851

    Sangroula, U., Han, K.-H., Koo, K.-M., Gnawali, K., Yum, K.-T.: Optimization of water distribution networks using genetic algorithm based sop–wdn program. Water14(6), 851 (2022) https://doi.org/10.3390/w14060851

    work page doi:10.3390/w14060851 2022

  11. [12]

    Environmental Modelling & Software14(5), 437–446 (1999) 29 https://doi.org/10.1016/s1364-8152(98)00089-9

    Gupta, I., Gupta, A., Khanna, P.: Genetic algorithm for optimization of water distribution systems. Environmental Modelling & Software14(5), 437–446 (1999) 29 https://doi.org/10.1016/s1364-8152(98)00089-9

    work page doi:10.1016/s1364-8152(98)00089-9 1999

  12. [13]

    Transportation Research Part B: Methodological197, 103238 (2025) https://doi.org/10.1016/j.trb.2025.103238

    Liang, M., Xu, M., Wang, S.: A novel multi-objective evolutionary algorithm for transit network design and frequency-setting problem considering passen- gers’ choice behaviors under station congestion. Transportation Research Part B: Methodological197, 103238 (2025) https://doi.org/10.1016/j.trb.2025.103238

    work page doi:10.1016/j.trb.2025.103238 2025

  13. [14]

    arXiv (2026)

    Mikaberidze, G., Mikaberidze, B., Taylor, D.: GradNet: A Gradient-Based Frame- work for Optimal Network Science. arXiv (2026). https://doi.org/10.48550/ ARXIV.2603.09197 . https://arxiv.org/abs/2603.09197

    Pith/arXiv arXiv 2026

  14. [15]

    A case study, pp

    Cembrano, G., Brdy´ s, M., Quevedo, J., Coulbeck, B., Orr, C.H.: Optimization of a multi-reservoir water network using a conjugate gradient technique. A case study, pp. 987–999. Springer, ??? (2006). https://doi.org/10.1007/bfb0042283 . http://dx.doi.org/10.1007/BFb0042283

    work page doi:10.1007/bfb0042283 2006

  15. [16]

    Antaki, G.A.: Piping and Pipeline Engineering: Design, Construction, Mainte- nance, Integrity, and Repair, 1st edn., p. 564. CRC Press, ??? (2003)

    2003

  16. [17]

    Proceedings of the Design Society5, 2571–2579 (2025) https://doi.org/10.1017/pds.2025.10271

    Clay, J., Sha, Z.: Paradigmatic design thinking: how generative ai changes the role of human designers. Proceedings of the Design Society5, 2571–2579 (2025) https://doi.org/10.1017/pds.2025.10271

    work page doi:10.1017/pds.2025.10271 2025

  17. [18]

    Proceedings of the Design Society4, 1959–1968 (2024) https://doi.org/10.1017/pds.2024.198

    Chiarello, F., Barandoni, S., Majda ˇSkec, M., Fantoni, G.: Generative large lan- guage models in engineering design: opportunities and challenges. Proceedings of the Design Society4, 1959–1968 (2024) https://doi.org/10.1017/pds.2024.198

    work page doi:10.1017/pds.2024.198 1959

  18. [19]

    The International Journal of Advanced Manufacturing Technology (2025) https://doi.org/10.1007/s00170-025-15830-2

    Daareyni, A., Martikkala, A., Mokhtarian, H., Ituarte, I.F.: Generative ai meets cad: enhancing engineering design to manufacturing processes with large language models. The International Journal of Advanced Manufacturing Technology (2025) https://doi.org/10.1007/s00170-025-15830-2

    work page doi:10.1007/s00170-025-15830-2 2025

  19. [20]

    arXiv preprint arXiv:2303.18223 (2023)

    Zhao, W.X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al.: A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)

    Pith/arXiv arXiv 2023

  20. [21]

    arXiv (2022)

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., Zhou, D.: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv (2022). https://doi.org/10.48550/ARXIV.2201.11903 . https://arxiv.org/abs/2201.11903

    Pith/arXiv arXiv doi:10.48550/arxiv.2201.11903 2022

  21. [22]

    In: The Eleventh International Conference on Learning Representations (2023)

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

    2023

  22. [23]

    In: Advances in Neural Information Processing Systems (2023)

    Schick, T., Dwivedi-Yu, J., Dess` ı, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves 30 to use tools. In: Advances in Neural Information Processing Systems (2023)

    2023

  23. [24]

    In: Advances in Neural Information Processing Systems (2023)

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

    2023

  24. [25]

    arXiv (2023)

    Wu, Y., Min, S.Y., Bisk, Y., Salakhutdinov, R., Azaria, A., Li, Y., Mitchell, T., Prabhumoye, S.: Plan, Eliminate, and Track – Language Models are Good Teachers for Embodied Agents. arXiv (2023). https://doi.org/10.48550/ARXIV. 2305.02412 . https://arxiv.org/abs/2305.02412

    Pith/arXiv arXiv doi:10.48550/arxiv 2023

  25. [26]

    In: Advances in Neural Information Processing Systems (2023)

    Li, G., Hammoud, H.A.A.A., Itani, H., Khizbullin, D., Ghanem, B.: Camel: Com- municative agents for” mind” exploration of large language model society. In: Advances in Neural Information Processing Systems (2023)

    2023

  26. [27]

    acl-long.810

    Qian, C., Liu, W., Liu, H., Chen, N., Dang, Y., Li, J., Yang, C., Chen, W., Su, Y., Cong, X., Xu, J., Li, D., Liu, Z., Sun, M.: Chatdev: Communicative agents for software development, 15174–15186 (2024) https://doi.org/10.18653/v1/2024. acl-long.810

    doi:10.18653/v1/2024 2024

  27. [28]

    arXiv preprint arXiv:2308.08155 (2023)

    Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J., et al.: Autogen: Enabling next-gen llm applications via multi-agent conversation. arXiv preprint arXiv:2308.08155 (2023)

    Pith/arXiv arXiv 2023

  28. [29]

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

    Hong, S., Zhuge, M., Chen, J., Zheng, X., Cheng, Y., Zhang, C., Wang, J., Wang, Z., Yau, S.S., Lin, Z.,et al.: Metagpt: Meta programming for a multi-agent collaborative framework. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  29. [30]

    Nature Computational Science 6(3), 301–315 (2025) https://doi.org/10.1038/s43588-025-00906-6

    Shao, E., Wang, Y., Qian, Y., Pan, Z., Liu, H., Wang, D.: Sciscigpt: advancing human–ai collaboration in the science of science. Nature Computational Science 6(3), 301–315 (2025) https://doi.org/10.1038/s43588-025-00906-6

    doi:10.1038/s43588-025-00906-6 2025

  30. [31]

    In: Advances in Neural Information Processing Systems, vol

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., K¨ uttler, H., Lewis, M., Yih, W.-t., Rockt¨ aschel, T.,et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. In: Advances in Neural Information Processing Systems, vol. 33, pp. 9459–9474 (2020)

    2020

  31. [32]

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

    Asai, A., Wu, Z., Wang, Y., Sil, A., Hajishirzi, H.: Self-rag: Learning to retrieve, generate, and critique through self-reflection. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  32. [33]

    arXiv preprint arXiv:2404.16130 (2024)

    Edge, D., Trinh, H., Cheng, N., Bradley, J., Chao, A., Mody, A., Truitt, S., Larson, J.: From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130 (2024)

    Pith/arXiv arXiv 2024

  33. [34]

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

    Ma, Y.J., Liang, W., Wang, G., Huang, D.-A., Bastani, O., Jayaraman, D., Zhu, 31 Y., Lin, L., Anandkumar, A.: Eureka: Human-level reward design via coding large language models. In: The Twelfth International Conference on Learning Representations (2024)

    2024

  34. [35]

    arXiv (2025)

    Xu, H., Wu, D., Chinazzi, M., Niu, R., Yu, R., Ma, Y.-A.: SimulRAG: Simulator- based RAG for Grounding LLMs in Long-form Scientific QA. arXiv (2025). https: //doi.org/10.48550/ARXIV.2509.25459 . https://arxiv.org/abs/2509.25459

    doi:10.48550/arxiv.2509.25459 2025

  35. [36]

    Expert Systems with Applications 314, 131505 (2026) https://doi.org/10.1016/j.eswa.2026.131505

    Venkatesh, P.R., P, R.K.: Ontollm: Enhancing llm grounding and digression pre- vention with ontologies and knowledge graphs. Expert Systems with Applications 314, 131505 (2026) https://doi.org/10.1016/j.eswa.2026.131505

    work page doi:10.1016/j.eswa.2026.131505 2026

  36. [37]

    arXiv (2024)

    Kommineni, V.K., K¨ onig-Ries, B., Samuel, S.: From human experts to machines: An LLM supported approach to ontology and knowledge graph construction. arXiv (2024). https://doi.org/10.48550/ARXIV.2403.08345 . https://arxiv.org/ abs/2403.08345

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.08345 2024

  37. [38]

    Liu, R., Luobei, L., Li, J., Wang, B., Liu, M., Wu, D., Wang, S., Qin, B.: Ontology-guided reverse thinking makes large language models stronger on knowl- edge graph question answering, 15269–15284 (2025) https://doi.org/10.18653/ v1/2025.acl-long.741

    2025

  38. [39]

    Nature Communications (2026) https:// doi.org/10.1038/s41467-026-71411-1

    Feng, Y., Hu, H., Ying, S., Hou, X., Liu, S., Yang, M., Li, J., Du, S., Zheng, N., Hu, H., Gao, Y.: Hyper-rag: combating llm hallucinations using hypergraph- driven retrieval-augmented generation. Nature Communications (2026) https:// doi.org/10.1038/s41467-026-71411-1

    doi:10.1038/s41467-026-71411-1 2026

  39. [40]

    IEEE Transactions on Power Delivery4(2), 1401– 1407 (1989) https://doi.org/10.1109/61.25627

    Baran, M.E., Wu, F.F.: Network reconfiguration in distribution systems for loss reduction and load balancing. IEEE Transactions on Power Delivery4(2), 1401– 1407 (1989) https://doi.org/10.1109/61.25627

    doi:10.1109/61.25627 1989

  40. [41]

    Water Resources Research13(6), 885–900 (1977) https://doi.org/10.1029/ wr013i006p00885

    Alperovits, E., Shamir, U.: Design of optimal water distribution systems. Water Resources Research13(6), 885–900 (1977) https://doi.org/10.1029/ wr013i006p00885

    1977

  41. [42]

    next speaker

    Savic, D.A., Walters, G.A.: Genetic algorithms for least-cost design of water distri- bution networks. Journal of Water Resources Planning and Management123(2), 67–77 (1997) https://doi.org/10.1061/(asce)0733-9496(1997)123:2(67) 32 A Agent Prompts & Workflows A.1 Z-COPA Agent Prompts This appendix presents compact prompt cards for the Z-COPA multi-agent w...

    work page doi:10.1061/(asce)0733-9496(1997)123:2(67 1997