Pith. sign in

REVIEW 6 major objections 5 minor 43 references

Entropy-Constrained Strategy Optimization in Urban Floods: A Multi-Agent Framework with LLM and Knowledge Graph Integration

T0 review · 6 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A hierarchical LLM framework with entropy-constrained generation outperforms rule-based and reinforcement-learning dispatch in urban flood simulations.

desk verdict The central claim runs ahead of the evidence: H-J's own results show PPO beating it on the task success metric in one of three scenarios, and the entropy-control mechanism is never tied to the claimed policy-level constraints. read the letter →

arxiv 2508.14654 v1 pith:3P7WNR64 submitted 2025-08-20 cs.AI

Peilin Ji , Xiao Xue , Simeng Wang , Wenhao Yan This is my paper
classification cs.AI
keywords urbanfloodemergencydispatchmulti-agentLLMframeworkentropy-constrainedgenerationknowledgegraphretrievalclosed-loopfeedbackoptimizationagent-basedsimulationsemanticstability
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

The paper introduces H-J, a closed-loop, hierarchical multi-agent framework for urban flood emergency dispatch built around an LLM. Its claim is that adding dual-channel knowledge retrieval, entropy-constrained generation, and objective-driven feedback lets LLM-generated strategies stay stable enough to beat both static rules and a PPO agent on traffic smoothness, trip success rate, cancellation rate, and robustness under extreme, intermittent, and light rainfall. The three mechanisms are treated as complementary: knowledge indexing anchors the prompt, entropy control stabilizes and diversifies output, and feedback replans when the global objective degrades. If the claim holds, it would make LLM-based dispatch a plausible alternative to rule-based and RL systems for fast-changing disaster response.

What carries the argument

The load-bearing object is the entropy-constrained hierarchical policy. High-level planning entropy is held below a threshold tau by a loss that penalizes |H(rho_theta) - tau|, and local execution policies are constrained to conditional entropy no greater than the global policy's entropy, so that uncertainty propagates downward without growing. Around that constraint sits a closed loop: dual-channel knowledge retrieval (GraphSAGE on a flood knowledge graph plus semantic retrieval over text logs) grounds the prompt; a weighted macro-objective J = omega_1 f + omega_2 t + omega_3 c + omega_4(1 - r) evaluates the outcome; and when J drops delta below its historical best, the system regenerates t

What would settle it

Measure the strategy-level entropy H(pi_h) from Eq. (3) and the conditional entropy H(pi_l | pi_h) under full H-J and under the no-entropy-control ablation on identical prompts; if the two are nearly equal, the entropy constraint is not the active mechanism and the ablation gains must be explained by something else. A complementary check is to rerun the three scenarios on a second city's rainfall record and road network and see whether H-J still beats PPO.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that an LLM dispatch system can be made reliable in flood response by explicitly bounding the entropy of generated strategies. The paper formulates a global policy entropy H(pi_h), requires local policies to respect H(pi_l | pi_h) <= H(pi_h) <= tau, and enforces the bound through an entropy-regularized loss with an adaptive penalty. In simulations on a real urban topology under three rainfall regimes, full H-J shows the most stable objective trajectory, the best overall success rate and cancellation rate, and lower run-to-run variance than the alternatives; ablations attribute the largest drop to removing dual-channel knowledge indexing, the largest volatil

Load-bearing premise

The framework's advantage rests on the assumption that the entropy-regularized loss applied at the token level actually constrains the strategy-level entropy in the paper's equations, and that the agent-based simulation with synthetic rainfall is faithful enough that beating PPO and rules there will transfer to real floods.

Editorial extensions

If this is right

  • In all three rainfall scenarios, H-J keeps the global objective J more stable than PPO, rules, or no planning, and produces lower cancellation and higher success rates on average.
  • Removing dual-channel knowledge indexing causes the largest performance decline, especially under extreme rainfall, so prompt grounding carries much of the system's robustness.
  • Removing entropy control raises variance and cuts semantic diversity (SDS from 0.443 to 0.312), so the entropy term is what the paper credits for balancing stable and diverse strategy output.
  • Removing feedback raises variance by about 20 percent, implying long-horizon adaptation in this setup comes mainly from the replanning trigger.
  • Because H-J is risk-averse and PPO can win on short-term arrival rate in intermittent bursts, the practical choice between the two depends on whether one values peak short-term throughput or stable global performance.

Reading between the lines

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

  • The paper never derives a link between the token-level entropy loss of Eq. (6) and the strategy-level entropy of Eqs. (3)-(5); the natural test is to measure H(pi_h) and the conditional entropy directly before and after fine-tuning.
  • A transfer test the paper does not run: apply the same H-J pipeline to a different urban flood topology or to another LLM-based dispatch domain (wildfire evacuation, hospital bed allocation) and check whether SCS/SDS and execution variance move the same way.
  • If entropy control is doing the claimed work, its effect should survive changing the base LLM; if the benefit mostly comes from knowledge indexing, swapping the LLM should leave SCS roughly unchanged, which would separate the two mechanisms empirically.
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.

Referee Report

6 major / 5 minor

Summary. The paper proposes H–J, a hierarchical multi-agent framework for urban flood emergency dispatch that combines LLM-based strategy generation, dual-channel knowledge retrieval (knowledge graph + unstructured text), entropy-constrained hierarchical generation, and a feedback loop driven by a weighted macro-objective J. The framework is evaluated in a Mesa ABM simulator under three rainfall scenarios (extreme rainfall, intermittent bursts, daily light rain) against empty, rule-based, and PPO baselines, with additional ablations. The central claim, stated in the abstract and Section 4.4, is that H–J significantly outperforms both rule-based and RL baselines in traffic smoothness, task success rate, and system robustness.

Significance. If the empirical claims were fully supported, this would be a useful contribution to LLM-based multi-agent decision support for disaster response: the closed-loop generation–execution–feedback architecture is coherent, the integration of knowledge-graph retrieval with entropy control is a sensible engineering direction, and the use of a real urban topology is a strength. The ablation design targets the three main components. However, the paper's central claims are undermined by an internal contradiction in the reported results, missing derivations connecting the entropy regularizer to the stated policy-level constraints, and an evaluation protocol that cannot support 'significantly outperforms' claims. As presented, the results do not substantiate the headline conclusions.

major comments (6)
  1. [Abstract / §4.4.1] The abstract claims H–J 'significantly outperforms both rule-based and reinforcement learning baselines in ... task success rate.' Yet §4.4.1 states that in the intermittent-burst scenario 'PPO slightly surpasses H–J in arrival rate.' Arrival rate r is defined in §4.1.3 and Appendix D.5 as the successful arrival ratio, i.e., the task-success metric used in the objective J. No numeric table, confidence interval, or significance test accompanies this comparison. This is an internal contradiction of the central claim, not merely a question of statistical strength. The claim must be reconciled or the abstract and conclusion revised.
  2. [§3.3.1 / §3.3.2 / Eq. (6)] The entropy-constrained loss in Eq. (6) is written as L = E[log ρθ(π)] − λ|H(ρθ) − τ|. No derivation or formal argument shows that this token-level entropy penalty bounds or controls the strategy-level entropies H(πh) and H(πl|πh) in Eqs. (3)–(5). In addition, the LoRA fine-tuning data are never described: which prompts, which outputs, what supervision, and how Eq. (6) is optimized. The paper therefore does not establish the central mechanism by which entropy-constrained generation stabilizes strategy-level outputs.
  3. [§4.1.3 / Table 2 / §4.5.1] The semantic evaluation is circular with respect to the intervention. SCS measures the average similarity of multiple responses to the same prompt, and the entropy-constrained loss is explicitly designed to make generation more deterministic; higher SCS under 'w/o Entropy Control' vs. full H–J is therefore expected by construction. Likewise SDS measures embedding spread, which the entropy/sampling mechanism directly shapes. These metrics do not measure semantic stability or execution consistency at the strategy level. The paper needs independent metrics, e.g., grounded instruction success or external human eval, before claiming entropy control improves 'strategy diversity' and 'semantic consistency'.
  4. [§4.2 / Table 1 / §4.4.1] No statistical test, confidence interval, or error bar is reported for the macroscopic metrics. The text says each strategy was repeated at least five times and mean/variance were recorded, but no numerical values are given for Figure 4 or the comparative claims. The phrase 'significantly outperforms' is unsupported. Moreover, Table 1 notes that all hyperparameters were tuned via preliminary runs on the same scenarios, and the baselines are not described in enough detail (e.g., PPO state space, reward, training time) to assess fairness. This is a load-bearing evaluation gap.
  5. [§3.4.1 vs. Appendix D.2 / Table 1] There is an internal inconsistency about the macro-objective weights. Section 3.4.1 states that 'the weights ωi are adapted via experience replay and short-term performance tracking,' while Appendix D.2 states that 'we adopt a fixed set of weights ωi' determined by grid search, and Table 1 lists a fixed weight vector. Similarly, Eq. (10) uses λ as a coefficient in the adaptive trigger while λ is earlier defined as the entropy penalty coefficient in Eqs. (6)–(7). These inconsistencies need to be resolved because the feedback-adaptation claim is a core part of the framework.
  6. [§4.1.1 / Appendix A.1 / B.1] The abstract and §4.1.1 state the evaluation uses 'real-world urban topology and rainfall data,' but Appendix A.1 says rainfall curves are 'synthesized' and Appendix B.1 says 'synthetic rainfall patterns are generated from historical meteorological statistics.' This is material: the scenario-stress realism affects the external-validity claim. The paper should state clearly which components are real and which are synthetic, and discuss the implications for transfer to practice.
minor comments (5)
  1. [Notation / Table 1] Eq. (8) uses ωi for objective weights, while Table 1 uses w; Eq. (10) overloads λ with the entropy penalty coefficient from Eqs. (6)–(7). Please use distinct symbols.
  2. [§4.2] The baseline 'Ruled' is presumably 'Rule-based'; use consistent terminology throughout.
  3. [Table 2] The 'Stability' metric is not defined precisely; it is described as variance of f, t, c, r in §4.1.3, but the reported value 0.0047 has no unit or aggregation formula. Please give the exact computation.
  4. [Eq. (11)] The deviation metric ∆E is defined but never used in any experiment or ablation. Either use it or remove it.
  5. [Reproducibility] No code, data, or trained-model release is mentioned. Given the number of tunable components, providing an artifact would substantially strengthen reproducibility.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-referential semantic metric; core traffic/task comparisons are not circular.

  1. self definitional [§4.5.1 / Table 2; Appendix D.6; §3.3.2 Eq. (6)]
    "Notably, entropy control raises SDS from 0.312 to 0.443 and improves SCS by 0.021, revealing its role in balancing diversity and coherence."

    SCS is defined in Appendix D.6 as the average pairwise embedding similarity of responses to the same prompt. Equation (6) explicitly minimizes |H(ρθ)-τ|, i.e., it suppresses the entropy of the model's output distribution. Lower output entropy mechanically makes repeated samples more similar, so the reported SCS gain under entropy control is a direct consequence of the objective being optimized. This is a self-consistency check rather than an independent validation of semantic consistency. It is not load-bearing for the headline comparisons on J, traffic flow, cancellation, or arrival rate, which are empirical simulation results.

full rationale

The central claimed result—H-J outperforming rule-based and PPO baselines on J, traffic flow, cancellation rate, and arrival rate—rests on simulation experiments, not on a derivation from the method's own definitions. The only place where the argument approaches circularity is the semantic metric SCS: Eq. (6) optimizes |H(ρθ)-τ| and Appendix D.6 defines SCS as same-prompt output similarity, so improving SCS under entropy control is a by-construction property rather than independent evidence. This metric supports the 'semantic consistency' portion of the conclusions but is not the headline traffic/task-success claim. No load-bearing self-citation chains or imported uniqueness theorems are present; the cited CPSS background [20-28] is contextual. The reported PPO advantage in arrival rate in §4.4.1 is an internal consistency issue, not a circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The framework's headline outputs rest on many hand-tuned numbers, most notably w, τ, α, δ, and the simulation scale, all selected on the same scenarios tested. The theoretical assumptions (token-to-policy entropy equivalence, semantic similarity as embedding cosine) are asserted, not derived. No independent benchmarks or external data are used.

free parameters (7)
  • Entropy threshold τ = 1.2
    Used in Eq. (5) and Eq. (6); chosen by hand in Table 1.
  • Initial penalty coefficient λ = 1.0 (dynamic)
    Start value for entropy penalty in Eq. (6); updated via Eq. (7).
  • Rhythm modulation rate α = 0.05
    Learning rate for λ update in Eq. (7).
  • Objective weights ω = [0.3, 0.3, 0.2, 0.2]
    Weights in Eq. (8); Table 1 gives fixed vector, Appendix D.2 says grid-searched.
  • Trigger threshold δ = 0.015
    Feedback replanning threshold in Eq. (10)/Table 1.
  • Feedback window length = 10
    Sliding window for J statistics in Eq. (10).
  • Region partitions / agent population / perception radius = 64 / 500 / 3
    Simulation configuration; affect the difficulty of dispatch.
assumptions (5)
  • domain assumption The Mesa ABM simulation with synthesized rainfall curves is a faithful proxy for real urban flood dynamics.
    Used in Section 4.1; the empirical claim is only as strong as the simulator.
  • ad hoc to paper Token-level entropy of LLM outputs corresponds to strategy-level policy entropy in Eqs. (3)-(5).
    §3.3.1 and §3.3.2 equate H(π_h) over actions with ρ_θ over tokens without defining the mapping.
  • domain assumption Sentence-BERT cosine similarity captures semantic consistency and diversity for LLM strategies.
    Appendix D.6-D.7; treats embedding similarity as semantic equivalence without validation in this domain.
  • domain assumption Weighted sum J in Eq. (8) is an adequate scalar objective for emergency dispatch quality.
    §3.4.1 and D.2; the paper acknowledges Pareto alternatives but does not validate the weights.
  • domain assumption The zero-shot instruction classifier reliably maps natural language strategies to executable commands.
    Appendix C.2; no accuracy measurement is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Entropy-Constrained Strategy Optimization in Urban Floods: A Multi-Agent Framework with LLM and Knowledge Graph Integration." pith.science (2026). https://pith.science/paper/3P7WNR64

@misc{pith2026250814654,
  author       = {Pith},
  title        = {Pith review of: Entropy-Constrained Strategy Optimization in Urban Floods: A Multi-Agent Framework with LLM and Knowledge Graph Integration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3P7WNR64}},
  note         = {Machine review of arXiv:2508.14654}
}
read the original abstract

In recent years, the increasing frequency of extreme urban rainfall events has posed significant challenges to emergency scheduling systems. Urban flooding often leads to severe traffic congestion and service disruptions, threatening public safety and mobility. However, effective decision making remains hindered by three key challenges: (1) managing trade-offs among competing goals (e.g., traffic flow, task completion, and risk mitigation) requires dynamic, context-aware strategies; (2) rapidly evolving environmental conditions render static rules inadequate; and (3) LLM-generated strategies frequently suffer from semantic instability and execution inconsistency. Existing methods fail to align perception, global optimization, and multi-agent coordination within a unified framework. To tackle these challenges, we introduce H-J, a hierarchical multi-agent framework that integrates knowledge-guided prompting, entropy-constrained generation, and feedback-driven optimization. The framework establishes a closed-loop pipeline spanning from multi-source perception to strategic execution and continuous refinement. We evaluate H-J on real-world urban topology and rainfall data under three representative conditions: extreme rainfall, intermittent bursts, and daily light rain. Experiments show that H-J outperforms rule-based and reinforcement-learning baselines in traffic smoothness, task success rate, and system robustness. These findings highlight the promise of uncertainty-aware, knowledge-constrained LLM-based approaches for enhancing resilience in urban flood response.

Figures

Figures reproduced from arXiv: 2508.14654 by the authors.

Figure 1
Figure 1. System-level overview of the proposed H–J framework. The architecture integrates multi-source perception, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the H–J framework, which integrates: (1) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Human flow density heatmaps at representative steps (5, 30/40, 45). The H–J strategy dynamically shifts [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparative performance under three rainfall scenarios. H–J consistently outperforms baselines across key [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Ablation results under three rainfall scenarios. Removing any module causes performance drops, especially [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Normalized rainfall intensity trends for extreme, intermittent, and light rain scenarios. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages

  1. [1]

    Shortest path planning and dynamic rescue forces dispatching for urban flood disasters.Scientific Reports, 15:23643, 2025

    Jun Liang, Mingyu Liu, Zongjia Zhang, and Lili Yang. Shortest path planning and dynamic rescue forces dispatching for urban flood disasters.Scientific Reports, 15:23643, 2025. 11 APREPRINT- SEPTEMBER13, 2025

    work page 2025

  2. [2]

    Adaptive Reinforcement Learning Model for Simulation of Urban Mobility during Crises

    Chao Fan, Xiangqi Jiang, and Ali Mostafavi. Adaptive reinforcement learning model for simulation of urban mobility during crises.arXiv preprint arXiv:2009.01359, 2020

    work page Pith review arXiv 2009

  3. [3]

    A reinforcement learning-based routing algorithm for large-scale evacuations in urban disaster scenarios.Tunnelling and Underground Space Technology, 2024

    Dong Li et al. A reinforcement learning-based routing algorithm for large-scale evacuations in urban disaster scenarios.Tunnelling and Underground Space Technology, 2024

    work page 2024

  4. [4]

    Reinforcement learning for optimizing evacuation routes in urban disaster scenarios,

    Faith Victoria Emmanuel et al. Reinforcement learning for optimizing evacuation routes in urban disaster scenarios,

    work page

  5. [5]

    Real-time Spatial Retrieval Augmented Generation for Urban Environments

    David Nazareno Campo, Javier Conde, et al. Real-time spatial retrieval augmented generation for urban environ- ments.arXiv preprint arXiv:2505.02271, 2025

    work page Pith review arXiv 2025

  6. [6]

    Streetviewllm: Extracting geographic information using a chain-of-thought multimodal large language model.arXiv preprint arXiv:2411.14476, 2024

    Zongrong Li, Junhao Xu, et al. Streetviewllm: Extracting geographic information using a chain-of-thought multimodal large language model.arXiv preprint arXiv:2411.14476, 2024

    arXiv 2024

  7. [7]

    Deep reinforcement learning for adaptive traffic signal control

    Kai Liang Tan, Subhadipto Poddar, and Soumik Sarkar. Deep reinforcement learning for adaptive traffic signal control. InNeurIPS Workshop on Machine Learning for Intelligent Transportation, 2019

    work page 2019

  8. [8]

    Li et al

    M. Li et al. Federated deep reinforcement learning-based urban traffic signal control.IEEE Transactions on Intelligent Transportation Systems, 2025

    work page 2025

Show all 43 references
  1. [9]

    Adaptive traffic signal control based on multi-agent reinforcement learning: Case study

    Dickness Kwakita, Angshuman Guin, and Michael Hunter. Adaptive traffic signal control based on multi-agent reinforcement learning: Case study... InArXiv preprint, 2025

    2025

  2. [10]

    Miletic, E

    M. Miletic, E. Ivanjko, et al. A review of reinforcement learning applications in adaptive traffic signal control. IET Intelligent Transport Systems, 2022

    2022

  3. [11]

    Michailidis, I

    P. Michailidis, I. Michailidis, C. R. Lazaridis, and E. Kosmatopoulos. Traffic signal control via reinforcement learning: A review on applications and innovations.Infrastructures, 10(5):114, 2025

    2025

  4. [12]

    El-Tantawy, B

    S. El-Tantawy, B. Abdulhai, and H. Abdelgawad. Multiagent reinforcement learning for integrated network of adaptive traffic signal controllers (marlin-atsc).IEEE Transactions on Intelligent Transportation Systems, 14(3): 1140–1150, 2013

    2013

  5. [13]

    W. Du, J. Ye, et al. Safelight: A reinforcement learning method toward collision-free traffic signal control. In AAAI Conference on Artificial Intelligence, volume 37, pages 14801–14810, 2023

    2023

  6. [14]

    Adaptive traffic signal control: Deep reinforcement learning algorithm with experience replay and target network.arXiv preprint arXiv:1705.02755, 2017

    Juntao Gao, Yulong Shen, et al. Adaptive traffic signal control: Deep reinforcement learning algorithm with experience replay and target network.arXiv preprint arXiv:1705.02755, 2017

    2017 arXiv

  7. [15]

    Llm agents for smart city management: Enhancing decision support through multi-agent ai systems.Smart Cities, 8(1):19, 2025

    Anna Kalyuzhnaya et al. Llm agents for smart city management: Enhancing decision support through multi-agent ai systems.Smart Cities, 8(1):19, 2025

    2025

  8. [16]

    A rag-based multi-agent llm system for natural hazard decision-support (wildfiregpt)

    David Nazareno Campo et al. A rag-based multi-agent llm system for natural hazard decision-support (wildfiregpt). arXiv preprint arXiv:2402.07877, 2025

    2025 arXiv

  9. [17]

    Li et al

    X. Li et al. Large language model powered intelligent urban agents.arXiv preprint arXiv:2507.00914, 2025

    2025 arXiv

  10. [18]

    Zhao et al

    P. Zhao et al. Extensible hierarchical multi-agent reinforcement-learning model for traffic signal control (gmhm). Applied Sciences, 12(24):12783, 2022

    2022

  11. [19]

    Xu et al

    B. Xu et al. Hierarchically and cooperatively learning traffic signal control. InAAAI Conference on Artificial Intelligence, volume 35, pages 669–677, 2021

    2021

  12. [20]

    Compu- tational experiments: Past, present and perspective.Acta Automatica Sinica, 49(2):246–271, 2023

    Xiao Xue, Xiang-Ning Yu, De-Yu Zhou, Chao Peng, Xiao Wang, Zhang-Bing Zhou, and Fei-Yue Wang. Compu- tational experiments: Past, present and perspective.Acta Automatica Sinica, 49(2):246–271, 2023

    2023

  13. [21]

    Computational experiments: A new method for analyzing cyber-physical-social systems (CPSS).IEEE Transactions on Systems, Man and Cybernetics: Systems, 54(2):813–826, 2024

    Xiao Xue, Yifan Shen, Xiangning Yu, De-Yu Zhou, Xiao Wang, Gang Wang, and Fei-Yue Wang. Computational experiments: A new method for analyzing cyber-physical-social systems (CPSS).IEEE Transactions on Systems, Man and Cybernetics: Systems, 54(2):813–826, 2024

    2024

  14. [22]

    Computational experiments for complex social systems part I: The customization of computational model.IEEE Transactions on Computational Social Systems, 9(5):1330–1344, 2022

    Xiao Xue, Fang-Yi Chen, De-Yu Zhou, Xiao Wang, Min Lu, and Fei-Yue Wang. Computational experiments for complex social systems part I: The customization of computational model.IEEE Transactions on Computational Social Systems, 9(5):1330–1344, 2022

    2022

  15. [23]

    Computational experiments for complex social system part II: the evaluation of computational model.IEEE Transactions on Computational Social Systems, 9(4):1224–1236, 2022

    Min Lu, Shizhan Chen, Xiao Xue, Xiao Wang, Yufang Zhang, and Fei-Yue Wang. Computational experiments for complex social system part II: the evaluation of computational model.IEEE Transactions on Computational Social Systems, 9(4):1224–1236, 2022

    2022

  16. [24]

    Computational experiments for complex social systems, part III: the docking of domain models.IEEE Transactions on Computational Social Systems, 11(2):1766–1780, 2024

    Xiao Xue, Xiangning Yu, Deyu Zhou, Chao Peng, Xiao Wang, and Fei-Yue Wang. Computational experiments for complex social systems, part III: the docking of domain models.IEEE Transactions on Computational Social Systems, 11(2):1766–1780, 2024. 12 APREPRINT- SEPTEMBER13, 2025

    2024

  17. [25]

    Computational experiments for complex social systems: Integrated design of experiment system.IEEE/CAA Journal of Automatica Sinica, 11(5):1175–1189, 2024

    Xiao Xue, Xiangning Yu, Deyu Zhou, Xiao Wang, Chongke Bi, Shufang Wang, and Fei-Yue Wang. Computational experiments for complex social systems: Integrated design of experiment system.IEEE/CAA Journal of Automatica Sinica, 11(5):1175–1189, 2024

    2024

  18. [26]

    Computational experiments for complex social systems: Experiment design and generative explanation.IEEE/CAA Journal of Automatica Sinica, 11(4):1022–1038, 2024

    Xiao Xue, Deyu Zhou, Xiangning Yu, Gang Wang, Juanjuan Li, Xia Xie, Lizhen Cui, and Fei-Yue Wang. Computational experiments for complex social systems: Experiment design and generative explanation.IEEE/CAA Journal of Automatica Sinica, 11(4):1022–1038, 2024

    2024

  19. [27]

    Social learning evolution (SLE): Computational experiment-based modeling framework of social manufacturing.IEEE Transactions on Industrial Informatics, 15(6):3343–3355, 2019

    Xiao Xue, Shufang Wang, Lejun Zhang, Zhiyong Feng, and Yaodan Guo. Social learning evolution (SLE): Computational experiment-based modeling framework of social manufacturing.IEEE Transactions on Industrial Informatics, 15(6):3343–3355, 2019

    2019

  20. [28]

    SLE2: The improved social learning evolution model of cloud manufacturing service ecosystem.IEEE Transactions on Industrial Informatics, 18(12):9017–9026, 2022

    Deyu Zhou, Xiao Xue, and Zhangbing Zhou. SLE2: The improved social learning evolution model of cloud manufacturing service ecosystem.IEEE Transactions on Industrial Informatics, 18(12):9017–9026, 2022

    2022

  21. [29]

    Urban flooding digital twin system framework.Systems Science & Control Engineering (Open Access), 13(1):2460432, 2025

    Chenyu Ge and Shengfeng Qin. Urban flooding digital twin system framework.Systems Science & Control Engineering (Open Access), 13(1):2460432, 2025. doi: 10.1080/21642583.2025.2460432

    2025

  22. [30]

    Retrieval-augmented generation across heterogeneous knowledge

    Wenhao Yu. Retrieval-augmented generation across heterogeneous knowledge. InNAACL Student Research Workshop, pages 52–58, 2022

    2022

  23. [31]

    Kg-retriever: Efficient knowledge indexing for retrieval-augmented large language models.arXiv preprint arXiv:2412.05547, 2024

    Weijie Chen, Ting Bai, Jinbo Su, et al. Kg-retriever: Efficient knowledge indexing for retrieval-augmented large language models.arXiv preprint arXiv:2412.05547, 2024

    2024 arXiv

  24. [32]

    Kg-infused rag: Augmenting corpus-based rag with external knowledge graphs.arXiv preprint arXiv:2506.09542, 2025

    Dingjun Wu, Yukun Yan, Zhenghao Liu, et al. Kg-infused rag: Augmenting corpus-based rag with external knowledge graphs.arXiv preprint arXiv:2506.09542, 2025

    2025

  25. [33]

    Semantic uncertainty in advanced decoding methods for llm generation.arXiv preprint arXiv:2506.17296, 2025

    Darius Foodeei, Simin Fan, and Martin Jaggi. Semantic uncertainty in advanced decoding methods for llm generation.arXiv preprint arXiv:2506.17296, 2025

    2025 arXiv

  26. [34]

    Min-p sampling for creative and coherent llm outputs

    Nguyen Nhat Minh, Andrew Baker, Clement Neo, et al. Min-p sampling for creative and coherent llm outputs. In ICLR 2025 (Oral), 2025

    2025

  27. [35]

    Improving diversity in language models: When temperature fails, change the loss.ICML 2025 Poster / Proceedings, 2025

    Alexandre Verine, Florian Le Bronnec, Kunhao Zheng, et al. Improving diversity in language models: When temperature fails, change the loss.ICML 2025 Poster / Proceedings, 2025

    2025

  28. [36]

    Hypertree planning: Enhancing llm reasoning via hierarchical thinking.ICML 2025, 2025

    Runquan Gui, Zhihai Wang, Jie Wang, et al. Hypertree planning: Enhancing llm reasoning via hierarchical thinking.ICML 2025, 2025

    2025

  29. [37]

    Y . J. Kim et al. Reinforcement learning with mpc-synchronized reward structure for adaptive traffic control. Applied Sciences, 14(12):5213, 2024

    2024

  30. [38]

    M. N. Gomes Jr et al. Real-time regulation of detention ponds via feedback control.Journal of Hydrology, 2024

    2024

  31. [39]

    Fehér et al

    Á. Fehér et al. Path planning via reinforcement learning with closed-loop stability guarantees.Transportation Research Part C, 2025

    2025

  32. [40]

    Predicting question- answering performance of large language models through semantic consistency.Findings of ACL Workshops 2023: GEM, 2023

    Ella Rabinovich, Samuel Ackerman, Orna Raz, Eitan Farchi, and Ateret Anaby-Tavor. Predicting question- answering performance of large language models through semantic consistency.Findings of ACL Workshops 2023: GEM, 2023

    2023

  33. [41]

    Score: Systematic consistency and robustness evaluation for large language models

    Grigor Nalbandyan et al. Score: Systematic consistency and robustness evaluation for large language models. arXiv preprint arXiv:2503.00137, 2025

    2025 arXiv

  34. [42]

    Exploring semantic consistency and style diversity for domain generalized semantic segmentation.arXiv preprint arXiv:2412.12050, 2024

    Hongwei Niu, Linhuang Xie, Jianghang Lin, and Shengchuan Zhang. Exploring semantic consistency and style diversity for domain generalized semantic segmentation.arXiv preprint arXiv:2412.12050, 2024. A Simulation Environment Details To evaluate the adaptability of the H–J frame...

    2024 arXiv

  35. [2025]

    Preprint posted May 2025

    2025

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.