Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

LLM-Based Community Surveys for Operational Decision Making in Interconnected Utility Infrastructures

T0 review · 4 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 claims that LLM-generated community personas, combined with a hetero-functional graph of infrastructure dependencies, can produce stable disaster repair priority rankings that recover the full-data ranking from just 60 percent…

desk verdict A candid proof of concept for LLM persona surveys in repair prioritization; internally coherent but the headline recovery claim needs an external benchmark before it carries operational weight. read the letter →

arxiv 2507.13577 v1 pith:WWXBUOCS submitted 2025-07-17 cs.SI

Adaeze Okeukwu-Ogbonnaya , Rahul Amatapu , Jason Bergtold , George Amariucai This is my paper
classification cs.SI
keywords hetero-functionalgraphinfrastructurerepairprioritizationLLMpersonascommunitypreferencespreferencelearningchainizationdisasterresiliencepairwisecomparisons
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

This paper proposes a way to decide which damaged utility should be repaired first after a disaster, in the common situation where engineering dependencies leave several repair orders equally valid. It models power, water, and community services as a hetero-functional graph, which yields a partial order of repair tasks, then uses an LLM to simulate 200 diverse residents answering pairwise "which first?" questions to turn that partial order into a complete priority list. The paper reports that a neural comparator plus chainization recovers the full-data ranking with Kendall's tau above 0.7 using only 60 percent of pairwise comparisons, and that top priorities are stable across prompt rewording. The authors present this as a proof of concept and explicitly caveat that LLM personas have not been validated against real residents.

What carries the argument

The central machinery is a hetero-functional graph, a directed graph whose nodes are functionalities such as generate power, transport water, or consume water in a residential area, and whose edges are dependency relations. This graph imposes a partial order of repair tasks, but the bottom-layer consumption nodes do not depend on one another, leaving many pairs incomparable. The paper resolves this ambiguity by training a neural pairwise comparator on soft labels aggregated from LLM-persona responses, applying the chainization algorithm to add pseudo-pairs, and then ranking all nodes with PageRank to produce a total order.

What would settle it

Collect real residents' pairwise repair preferences in communities matched to the simulated scenarios and compare the resulting global ranking with the LLM-persona ranking; if Kendall's tau between them is low, say below 0.5, the claim that these priorities reflect community preferences fails.

Watch

Extended reading notes

Core claim

The paper claims that a repair-priority total order can be produced by combining technical dependency structure with synthetic community preferences. In the toy three-community power and water system, the hetero-functional graph fixes which functionalities must be repaired before others, and the LLM personas resolve the remaining incomparabilities. The authors report that the resulting ranking can be approximated from 60 percent of the pairwise comparisons with Kendall's tau above 0.7, that the same global ranking emerges for community-aware and community-unaware personas, and that the Social Vulnerability Score is the dominant factor in the personas' choices. They present this as a proof of concept and state plainly that the simulated preferences may not match how real people behave in stressful situations.

Load-bearing premise

The load-bearing premise is that LLM personas built from census attributes and disaster scenarios produce preferences that can stand in for real community members during a disaster; the paper itself notes this is unvalidated.

Editorial extensions

If this is right

  • Using only 60 percent of the pairwise comparisons, chainization with pseudo-pairs recovers the full-data ranking with Kendall's tau above 0.7, so survey effort can be nearly halved.
  • The final global ranking was identical for community-aware and community-unaware personas, showing that community identity did not drive prioritization in this scenario.
  • Top-ranked repairs stay stable across prompt variations, while lower-ranked items shift; removing the Social Vulnerability Score explanation gives the highest agreement with the original ranking, tau = 0.72.
  • The system ranks school and residential repairs highest and places the most vulnerable community near the top, indicating that vulnerability weighs more than affiliation in LLM choices.
  • The unified ranking can be produced pre-disaster and used for planning and cross-sector investment decisions, not only for immediate response.

Reading between the lines

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

  • If the central claim holds, the dominance of the Social Vulnerability Score in LLM choices suggests real residents may also weight vulnerability heavily, but this needs direct elicitation to confirm.
  • The 60-percent recovery result implies active learning could select the most informative pairwise comparisons and cut survey burden further than random sampling does.
  • The method's reliance on simulated personas means it could be repurposed to generate synthetic preferences for new communities from prior survey data, a direction the authors flag for future work.
  • A single cross-sector repair list suits planning and investment decisions, but field deployment would need to respect specialized crews for water, power, and transportation.
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

4 major / 5 minor

Summary. The paper proposes a pipeline for disaster repair prioritization: a hetero-functional graph (HFG) imposes a partial order on repair tasks; LLM-generated personas, built from census-like attributes, answer pairwise comparison questions; a neural comparator learns preferences from aggregated soft labels; and a chainization algorithm with pseudo-pairs converts partial data into a total order. Experiments examine the effect of community awareness, partial-data recovery, and prompt sensitivity. The headline claim is that using 60% of the pairwise comparisons yields Kendall tau above 0.7 against the full-data ranking, and that top priorities are stable across prompt variations.

Significance. If the claims were externally validated, the paper would offer a scalable and low-cost way to elicit community input for infrastructure repair prioritization, and the public release of code and data is a clear strength. The prompt-sensitivity analysis is a useful first step toward understanding how LLM-generated preferences depend on prompt design. However, the current significance is limited because the evaluation is self-referential and the proxy-survey premise is unvalidated; as it stands, the result is best read as an internal-consistency proof of concept rather than a demonstrated recovery of community preferences.

major comments (4)
  1. [§III-E and §IV-B] The ground truth used in the recovery experiment is not an external or human-generated ranking: it is the output of the same neural comparator trained on all 36 pairwise questions followed by PageRank, and the pseudo-pairs in §III-E are generated from the model's own preliminary ranking with hard labels [1.0, 0.0]. The reported Kendall tau above 0.7 therefore measures self-consistency of the pipeline rather than recovery of a stable or externally meaningful repair order. I ask for evaluation against held-out pairwise judgments (e.g., leave-one-question-out or a random-feature baseline such as ranking by SVS alone) and for an ablation that removes pseudo-pair fine-tuning.
  2. [§III-B and §V] The paper's central premise is that LLM-generated personas stand in for real community members during disasters, but no evidence is offered that the simulated preferences match any real population. The manuscript itself concedes in §V that "the preferences generated by the model may not fully match how real people behave." Given that every downstream ranking, recovery experiment, and sensitivity analysis is built on these simulated preferences, the operational conclusions in the Discussion ("reliable prioritization") require either validation against real survey data or a careful narrowing of the claims to "internally consistent LLM-generated preferences." The community-aware versus community-unaware comparison does not address this external validity gap.
  3. [§IV-B and Figure 6] The recovery curves report only mean Kendall's tau and Top-K overlap, averaged over 100 random trials, with no error bars, confidence intervals, or per-trial variability. With only 36 pairwise questions and a small feature space, the reader cannot assess whether the improvement at 60% is statistically meaningful or an artifact of a particular random split. Please report standard deviations or confidence intervals and, ideally, significance tests against a random-feature or majority-vote comparator.
  4. [§V] The Discussion states that "we do not ... run explicit ablations to isolate the impact of pseudo-pair fine-tuning." Without such an ablation, the improvement attributed to chainization with pseudo-pairs could be due simply to training on additional synthetic data derived from the model's own predictions. An ablation comparing (i) no pseudo-pairs, (ii) pseudo-pairs from a random ranking, and (iii) pseudo-pairs from the model's preliminary ranking is needed to support the specific mechanism claimed in the headline result.
minor comments (5)
  1. [§III-B] Persona generation uses GPT-3.5-turbo at temperature 0.9 and response generation at 0.7, but no seed or reproducibility details are given beyond the date and time; given the stochasticity of the API, please report whether the main results are stable across repeated generations.
  2. [§III-D] The edge direction in the preference graph is defined as "when option 1 is preferred over option 2, an edge (option2 → option1) is added"; since PageRank is used instead of topological sorting, please clarify how edge weights and directions translate into the final ranking and whether cycles are resolved in a way that respects the HFG partial order.
  3. [§IV-C and Table IV] The claim that "the very top priorities remain stable" is not fully supported by Table IV: the top item changes across variants (e.g., Power Residential in Community 2 is first under "Without SVS" and "Choice First"), and several Kendall tau values have p > 0.05. Please qualify the stability claim.
  4. [§IV-A] The community-aware versus community-unaware comparison shows identical final rankings, but only one disaster scenario is tested; the conclusion that community identity does not influence prioritization is therefore limited to this toy setting and should be stated as such.
  5. [§II] Reference [12], which identifies systematic biases in LLM personas, is directly relevant to the paper's external validity limitation; consider using it more prominently in the Discussion to frame the scope of the claims.

Circularity Check

2 steps flagged · score 5.0 of 10

The headline recovery claim is an internal-consistency loop: the 'ground-truth' full-data ranking and the pseudo-pair labels are both outputs of the same LLM-persona + comparator pipeline, so the >0.7 Kendall tau is self-referential rather than external validation.

  1. fitted input called prediction [Section IV-B (Accuracy with Limited Pairwise Comparisons) and Section V (Discussion)]
    "Then, we applied Kendall’s τ correlation between the final inferred ranking and the ranking obtained from the full dataset. ... These results show that we can recover good global rankings from limited preference data, which offers a practical way to reduce the burden on survey participants while maintaining decision quality."

    The full-data ranking is not an external ground truth: it is produced by the same neural pairwise comparator trained on all 36 LLM-persona questions, followed by PageRank on the model's predicted preference graph. The subset experiment trains another instance of that comparator on a sample of the same questions, then measures agreement with the full-data model's output. Both sides of the Kendall tau are generated by the same LLM-synthetic responses and the same feature-based architecture, so the reported 'recovery' is an internal self-consistency check, not a prediction of real community preferences.

  2. fitted input called prediction [Section III-E (Random Subset Sampling and Fine-Tuning with Pseudo-Pairs)]
    "Based on the recovered initial global rank, we generate synthetic pseudo-pairs. We do this by assuming that for any two facilities that have a clear ranking relationship (e.g., A > B), a pseudo-comparison is added with a hard label indicating definite preference,i.e., [1.0, 0.0]."

    Pseudo-pairs are constructed from the model's own preliminary ranking: pairs with A>B in that ranking are turned into hard training labels [1.0,0.0]. Thus the model is fine-tuned on its own prior outputs; any resulting ranking is partially a reflection of those self-generated labels rather than of additional independent evidence from the sampled comparisons. Since the evaluation target is also an output of the same model family, the experiment is a closed loop of self-confirmation. The paper's own limitation statement admits no ablation was run to isolate the effect of pseudo-pair fine-tuning, so the contribution of this self-referential step to the reported tau cannot be separated.

full rationale

The paper is transparent that this is a proof of concept and that LLM personas are proxies; I do not count that as circularity by itself. The circular component is confined to the evaluation of the recovery claim. In Section IV-B, the ranking 'obtained from the full dataset' is not a measured community preference order; it is the PageRank of the preference graph predicted by the neural comparator trained on all 36 aggregated LLM responses. The subset model is trained on a random sample of those same questions and then fine-tuned on pseudo-pairs whose hard labels are generated from the subset model's own preliminary ranking (Section III-E). Comparing the subset ranking to the full-data model ranking therefore measures self-consistency of one synthetic-data/model pipeline, not whether 60% of questions recovers an accurate or externally meaningful order. The Discussion's 'reliable prioritization' inference overreaches. However, the result is not forced by construction: the subset model genuinely does not see the held-out comparisons, and the Kendall tau could in principle have been low. The circularity is partial and is exacerbated by the admitted absence of external benchmarks, expert-heuristic comparisons, and pseudo-pair ablations. Other components (HFG partial order, prompt sensitivity analysis, community-aware vs unaware comparison) are not circular. No load-bearing self-citation chain or imported uniqueness theorem is present; the relevant citations to chainization [33] and HFG [8-10] are external. I assign score 5.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical or formal entities are postulated. The free parameters are generative and training settings; the critical unvalidated assumption is that LLM personas represent real communities, which the paper itself flags as future work.

free parameters (5)
  • Temperature for persona generation = 0.9
    Chosen to promote diversity in synthetic personas; affects the sampled attributes and all downstream preferences.
  • Temperature for response generation = 0.7
    Chosen to encourage stable and coherent pairwise preference responses; directly shapes the survey data.
  • Learning rate for comparator network = 1e-3
    Selected by six-fold cross-validation over 1e-3, 5e-4, and 1e-4; influences the trained comparator and final rankings.
  • Batch size = 8
    Chosen for training the pairwise comparator; a minor optimization choice.
  • Social Vulnerability Score (SVS) scale in prompts = Inconsistent: paper states 0-1 normalized, prompt template says 0-1, but simulation setup and sample response use…
    Conflicting scales could materially change which communities the LLM treats as most vulnerable; the authors themselves report a misread of 0.2 as high vulnerability.
assumptions (4)
  • domain assumption The hetero-functional graph correctly encodes the repair dependencies of the toy infrastructure
    Section III-A defines the partial order from the HFG; the graph itself is provided in the GitHub repository and is treated as ground truth for technical feasibility.
  • domain assumption LLM personas with census attributes produce preferences representative of real community members
    Section III-B.1 uses GPT-3.5-turbo to create personas and treats their pairwise choices as community preferences; the paper cites Li et al. [12] showing such personas deviate from real data, and no real-survey validation is included.
  • standard math Aggregating persona votes into soft labels and learning a comparator yields a meaningful social consensus ranking
    Section III-D follows standard preference-learning practice; this is a modeling assumption about how individual choices combine.
  • standard math PageRank scores provide a consistent total order from the learned pairwise preference graph
    Section III-D applies PageRank to a directed graph that may contain cycles; PageRank is a standard ranking algorithm but the equivalence to a consensus ranking is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-Based Community Surveys for Operational Decision Making in Interconnected Utility Infrastructures." pith.science (2026). https://pith.science/paper/WWXBUOCS

@misc{pith2026250713577,
  author       = {Pith},
  title        = {Pith review of: LLM-Based Community Surveys for Operational Decision Making in Interconnected Utility Infrastructures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WWXBUOCS}},
  note         = {Machine review of arXiv:2507.13577}
}
read the original abstract

We represent interdependent infrastructure systems and communities alike with a hetero-functional graph (HFG) that encodes the dependencies between functionalities. This graph naturally imposes a partial order of functionalities that can inform the sequence of repair decisions to be made during a disaster across affected communities. However, using such technical criteria alone provides limited guidance at the point where the functionalities directly impact the communities, since these can be repaired in any order without violating the system constraints. To address this gap and improve resilience, we integrate community preferences to refine this partial order from the HFG into a total order. Our strategy involves getting the communities' opinions on their preferred sequence for repair crews to address infrastructure issues, considering potential constraints on resources. Due to the delay and cost associated with real-world survey data, we utilize a Large Language Model (LLM) as a proxy survey tool. We use the LLM to craft distinct personas representing individuals, each with varied disaster experiences. We construct diverse disaster scenarios, and each simulated persona provides input on prioritizing infrastructure repair needs across various communities. Finally, we apply learning algorithms to generate a global order based on the aggregated responses from these LLM-generated personas.

Figures

Figures reproduced from arXiv: 2507.13577 by the authors.

Figure 1
Figure 1. The Integrated Infrastructure and Commu [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. A subset of the hetero-functional graph (HFG) showing dependencies across power and water [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Six-fold cross-validation curves for the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Community Preferences by Community￾Aware Personas [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 6
Figure 6. Figure 6: shows the trend of mean Kendall’s τ scores across different percentages of compar￾isons used. Results are averaged over 100 ran￾dom trials for each sampling fraction. We observe that as the percentage of available comparisons increases, the agreement between the predic…
Figure 7
Figure 7. Figure 7: Top-K overlap on Rankings obtained via Random SubSampling and Pseudo-Pairs Tau and Top-k overlaps. We observe that chang￾ing instruction format gives τ= 0.50(p=0.075) and rewording the question in the prompts yields τ= 0.44(p=0.12), rewording the reasoning gives τ= 0.3…
Figure 9
Figure 9. Figure 9: shows the original prompt we use. This prompt format differs from the “Choice Before Reasoning” prompt, where the persona makes a choice before reasoning, as seen in this example response 15 [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 11
Figure 11. Figure 11: shows the prompt for shifting tone [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]
Figure 14
Figure 14. Figure 14: Full specification of one synthetic persona [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LISTEN to Your Preferences: An LLM Framework for Multi-Objective Selection

    cs.CL 2025-10 unverdicted novelty 6.0 of 10

    LISTEN uses LLMs as zero-shot preference oracles, via iterative utility refinement (LISTEN-U) or tournament comparisons (LISTEN-T), to select preferred items from large multi-objective candidate sets.

Reference graph

Works this paper leans on

41 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [1]

    A multidisciplinary definition and evaluation of resilience: The role of social justice in defining resilience,

    N. Doorn, P. Gardoni, and C. Murphy, “A multidisciplinary definition and evaluation of resilience: The role of social justice in defining resilience,” Sustainable and Resilient In- frastructure, vol. 4, no. 3, pp. 112–123, 2019

    work page 2019

  2. [2]

    Disaster planning and risk communication with vulnerable communities: lessons from hurricane katrina,

    D. P. Eisenman, K. M. Cordasco, S. Asch, J. F. Golden, and D. Glik, “Disaster planning and risk communication with vulnerable communities: lessons from hurricane katrina,” American journal of public health , vol. 97, no. Supplement_1, pp. S109–S115, 2007

    work page 2007

  3. [3]

    Disasters and communities: vul- nerability, resilience and preparedness,

    D. Paton and D. Johnston, “Disasters and communities: vul- nerability, resilience and preparedness,” Disaster Prevention and Management: An International Journal , vol. 10, no. 4, pp. 270–277, 2001

    work page 2001

  4. [4]

    Examine the effects of neighborhood equity on disaster situational awareness: Harness machine learning and geotagged twitter data,

    W. Zhai, Z.-R. Peng, and F. Yuan, “Examine the effects of neighborhood equity on disaster situational awareness: Harness machine learning and geotagged twitter data,” In- ternational Journal of Disaster Risk Reduction , vol. 48, p. 101611, 2020

    work page 2020

  5. [5]

    Social vulnera- bility and short-term disaster assistance in the united states,

    O. Drakes, E. Tate, J. Rainey, and S. Brody, “Social vulnera- bility and short-term disaster assistance in the united states,” International Journal of Disaster Risk Reduction , vol. 53, p. 102010, 2021

    work page 2021

  6. [6]

    Social vulnerability and procedural equity: Exploring the distribution of disaster aid across counties in the united states,

    S. J. Domingue and C. T. Emrich, “Social vulnerability and procedural equity: Exploring the distribution of disaster aid across counties in the united states,” The American Review of Public Administration , vol. 49, no. 8, pp. 897–913, 2019

    work page 2019

  7. [7]

    Public demand urgency for equitable infrastructure restoration planning,

    Y . Chen and W. Ji, “Public demand urgency for equitable infrastructure restoration planning,” International Journal of Disaster Risk Reduction , vol. 64, p. 102510, 2021

    work page 2021

  8. [8]

    W. C. Schoonenberg, I. S. Khayal, and A. M. Farid, A Hetero- functional graph theory for modeling interdependent smart City infrastructure. Springer, 2019

    work page 2019

Show all 41 references
  1. [9]

    A hetero-functional graph structural analysis of the american multi-modal energy sys- tem,

    D. J. Thompson and A. M. Farid, “A hetero-functional graph structural analysis of the american multi-modal energy sys- tem,” Sustainable Energy, Grids and Networks , vol. 38, p. 101254, 2024

    2024

  2. [10]

    Robustness assess- ment of hetero-functional graph theory based model of inter- dependent urban utility networks,

    S. Munikoti, K. Lai, and B. Natarajan, “Robustness assess- ment of hetero-functional graph theory based model of inter- dependent urban utility networks,” Reliability Engineering & System Safety , vol. 212, p. 107627, 2021

    2021

  3. [11]

    Agentic llm framework for adaptive decision discourse,

    A. Dolant and P. Kumar, “Agentic llm framework for adaptive decision discourse,” arXiv preprint arXiv:2502.10978 , 2025

    2025

  4. [12]

    Llm gen- erated persona is a promise with a catch,

    A. Li, H. Chen, H. Namkoong, and T. Peng, “Llm gen- erated persona is a promise with a catch,” arXiv preprint arXiv:2503.16527, 2025

    2025 arXiv

  5. [13]

    Argumentative experience: Reducing confir- mation bias on controversial issues through llm-generated multi-persona debates,

    L. Shi, H. Liu, Y . Wong, U. Mujumdar, D. Zhang, J. Gwizdka, and M. Lease, “Argumentative experience: Reducing confir- mation bias on controversial issues through llm-generated multi-persona debates,” arXiv preprint arXiv:2412.04629 , 2024

    2024

  6. [14]

    A rag-based multi-agent llm system for natural hazard resilience and adaptation,

    Y . Xie, B. Jiang, T. Mallick, J. D. Bergerson, J. K. Hutchison, D. R. Verner, J. Branham, M. R. Alexander, R. B. Ross, Y . Feng, L.-A. Levy, W. Su, and C. J. Taylor, “A rag-based multi-agent llm system for natural hazard resilience and adaptation,” 2025. [Online]. Available: ...

    2025 arXiv

  7. [15]

    Integrating sentiment analysis and reinforce- ment learning for equitable disaster response: A novel ap- proach

    S. Alqithami, “Integrating sentiment analysis and reinforce- ment learning for equitable disaster response: A novel ap- proach.” Sustainability (2071-1050) , vol. 17, no. 3, 2025

    2025

  8. [16]

    Intelligent urban emergency response: Integrating large language models, multi-objective optimization, and reinforcement learning,

    S. Chen, G. Luo, and L. Chen, “Intelligent urban emergency response: Integrating large language models, multi-objective optimization, and reinforcement learning,” in 2024 IEEE Smart World Congress (SWC) . IEEE, 2024, pp. 1469–1476

    2024

  9. [17]

    En- hancing emergency decision-making with knowledge graphs and large language models,

    M. Chen, Z. Tao, W. Tang, T. Qin, R. Yang, and C. Zhu, “En- hancing emergency decision-making with knowledge graphs and large language models,” International Journal of Disaster Risk Reduction , vol. 113, p. 104804, 2024

    2024

  10. [18]

    Eq- uitable prioritization of bridge rehabilitation projects using a spatial multi-criteria decision support framework,

    N. Mohamadiazar, A. Ebrahimian, and A. Azizinamini, “Eq- uitable prioritization of bridge rehabilitation projects using a spatial multi-criteria decision support framework,” Trans- portation Research Record , p. 03611981241254110, 2024

    2024

  11. [19]

    A decision support system for infrastructure rehabilitation planning,

    M. Hastak, Q. Cui, B. Safi, and S. Gokhale, “A decision support system for infrastructure rehabilitation planning,” in Computing in Civil Engineering (2005) , 2005, pp. 1–10

    2005

  12. [20]

    A multi-criteria decision-making model using socio-technical attributes for transportation bridge maintenance prioritisation,

    R. Das and M. Nakano, “A multi-criteria decision-making model using socio-technical attributes for transportation bridge maintenance prioritisation,” International Journal of Construction Management, vol. 23, no. 4, pp. 579–585, 2023

    2023

  13. [21]

    Decision-making criteria and indicators for water and sanitation projects in developing countries,

    M. Garfì and L. Ferrer-Martí, “Decision-making criteria and indicators for water and sanitation projects in developing countries,” Water Science and Technology , vol. 64, no. 1, pp. 83–101, 2011

    2011

  14. [22]

    Multicriteria-decision support for master plan scheduling: urban regeneration of an industrial area in north- ern italy,

    F. Dell’Anna, M. Berta, M. Bottero, G. Mallia, and V . Morgese, “Multicriteria-decision support for master plan scheduling: urban regeneration of an industrial area in north- ern italy,” Construction Management and Economics , vol. 42, no. 5, pp. 476–501, 2024

    2024

  15. [23]

    Prioritizing infrastructure investment: a framework for gov- ernment decision making,

    D. Marcelo, C. Mandri-Perrott, S. House, and J. Schwartz, “Prioritizing infrastructure investment: a framework for gov- ernment decision making,” World Bank Policy Research Work- ing Paper, no. 7674, 2016

    2016

  16. [24]

    Design of a prioritization methodology for equitable infrastructure planning,

    R. Dhansinghani, A. Ibrahim, A. Kannoth, C. Miller, L. Nguyen, S. Pham, and R. Bailey, “Design of a prioritization methodology for equitable infrastructure planning,” in 2022 Systems and Information Engineering Design Symposium (SIEDS). IEEE, 2022, pp. 315–320

    2022

  17. [25]

    Priority-based decision model for rehabilitation of water networks using fahp,

    S. Karasneh and S. Moqbel, “Priority-based decision model for rehabilitation of water networks using fahp,” Water Con- servation & Management , vol. 8, no. 1, 2024

    2024

  18. [26]

    Integrating sustainability in municipal wastew- ater infrastructure decision-analysis using the analytic hi- erarchy process,

    H. S. Galal, “Integrating sustainability in municipal wastew- ater infrastructure decision-analysis using the analytic hi- erarchy process,” Ph.D. dissertation, University of British Columbia, 2013

    2013

  19. [27]

    The bridge maintenance factors model: a pls-sem approach,

    N. P. Pramesti, H. Priyosulistyo, A. Aminullah, and A. Koes- margono, “The bridge maintenance factors model: a pls-sem approach,” Age, vol. 22, p. 18, 2021

    2021

  20. [28]

    Multi-criteria decision support for annual rehabilitation programmes in drinking wa- ter networks,

    R. Baur, P. Le Gauffre, and S. Sægrov, “Multi-criteria decision support for annual rehabilitation programmes in drinking wa- ter networks,” Water Science and Technology: Water Supply , vol. 3, no. 1-2, pp. 43–50, 2003

    2003

  21. [29]

    Preference completion: Large-scale collaborative ranking from pairwise comparisons,

    D. Park, J. Neeman, J. Zhang, S. Sanghavi, and I. Dhillon, “Preference completion: Large-scale collaborative ranking from pairwise comparisons,” in International Conference on Machine Learning . PMLR, 2015, pp. 1907–1916

    2015

  22. [30]

    Preference completion from partial rankings,

    S. Gunasekar, O. O. Koyejo, and J. Ghosh, “Preference completion from partial rankings,” Advances in Neural In- formation Processing Systems , vol. 29, 2016

    2016

  23. [31]

    Semi-supervised learning to rank with preference regularization,

    M. Szummer and E. Yilmaz, “Semi-supervised learning to rank with preference regularization,” in Proceedings of the 20th ACM international conference on Information and knowledge management, 2011, pp. 269–278

    2011

  24. [32]

    Gnnrank: Learning global rankings from pairwise comparisons via directed graph neural networks,

    Y . He, Q. Gan, D. Wipf, G. D. Reinert, J. Yan, and M. Cu- curingu, “Gnnrank: Learning global rankings from pairwise comparisons via directed graph neural networks,” in interna- tional conference on machine learning . PMLR, 2022, pp. 8581–8612

    2022

  25. [33]

    Order learning using partially ordered data via chainization,

    S.-H. Lee and C.-S. Kim, “Order learning using partially ordered data via chainization,” in European Conference on Computer Vision. Springer, 2022, pp. 196–211

    2022

  26. [34]

    CommRank-HFG,

    A. Okeukwu-Ogbonnaya, “CommRank-HFG,” https://github. com/adaezy/CommRank-HFG, 2025

    2025

  27. [35]

    Social vulnerability score: a scalable index for representing social vulnerability in virtual community resilience testbeds,

    S. A. Enderami and E. Sutley, “Social vulnerability score: a scalable index for representing social vulnerability in virtual community resilience testbeds,” Natural Hazards , vol. 120, no. 7, pp. 6457–6480, 2024

    2024

  28. [36]

    Ranking under uncer- tainty,

    O. Zuk, L. Ein-Dor, and E. Domany, “Ranking under uncer- tainty,” arXiv preprint arXiv:1206.5280 , 2012

    2012 arXiv

  29. [37]

    Koios: Top-k semantic overlap set search,

    P. Mundra, J. Zhang, F. Nargesian, and N. Augsten, “Koios: Top-k semantic overlap set search,” in 2023 IEEE 39th Inter- national Conference on Data Engineering (ICDE) . IEEE, 2023, pp. 1531–1543

    2023

  30. [38]

    State of what art? a call for multi- prompt llm evaluation,

    M. Mizrahi, G. Kaplan, D. Malkin, R. Dror, D. Shahaf, and G. Stanovsky, “State of what art? a call for multi- prompt llm evaluation,” Transactions of the Association for Computational Linguistics , vol. 12, pp. 933–949, 2024

    2024

  31. [39]

    Posix: A prompt sensitivity index for large language models,

    A. Chatterjee, H. K. Renduchintala, S. Bhatia, and T. Chakraborty, “Posix: A prompt sensitivity index for large language models,” arXiv preprint arXiv:2410.02185 , 2024

    2024 arXiv

  32. [40]

    Prompt sentiment: The catalyst for llm change,

    V . Gandhi and S. Gandhi, “Prompt sentiment: The catalyst for llm change,” arXiv preprint arXiv:2503.13510 , 2025

    2025 arXiv

  33. [41]

    Does prompt formatting have any impact on llm performance?

    J. He, M. Rungta, D. Koleczek, A. Sekhon, F. X. Wang, and S. Hasan, “Does prompt formatting have any impact on llm performance?” arXiv preprint arXiv:2411.10541 , 2024

    2024 arXiv

Pith tools

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