PaperlayerView exact revision on arXiv

Source-linked AI summary

LLM4Trust: Exploring the Capabilities of Large Language Models for Trust Evaluation

Jie Wang, Yanbo Sun, Zheng Yan, Jiahe Lan, Elisa Bertino

arXiv:2609.33521v2cs.LG

TL;DR

Learning-based trust evaluation often requires substantial labels and costly training, can miss basic trust properties, and offers limited explanations; existing datasets also lack explicit property annotations, while LLM context windows constrain graph inputs. LLM4Trust benchmarks eight LLMs across five trust properties and applies selected model-prompt combinations to real-world datasets using strategies to extract relevant graph information. The authors report effective property understanding and promise under limited supervision, alongside vulnerabilities to attacks and high inference costs.

  • Problem

    Learning-based trust evaluation often needs substantial ground-truth labels, costly training, and offers limited trust-property coverage and explanation; datasets also lack explicit property annotations.

    Source: §I · p. 2

  • Method

    LLM4Trust tests eight LLMs with nine prompt methods on five trust-property tasks, then evaluates selected combinations on real-world datasets using strategies to fit relevant graph information into prompts.

    Source: §I · p. 2

  • Results

    LLMs generally understand the five basic trust properties and show strong potential for real-world evaluation, particularly under limited supervision.

    Source: §I · p. 2, §VII · p. 14

  • Takeaways & Limitations

    LLMs may support trust evaluation where supervision is limited, with transparent inference processes that enhance explainability.

    Source: §VII · p. 14

    LLM evaluations take several seconds per trustor-trustee pair, compared with learning-based methods evaluating thousands of pairs in under one second.

    Source: §VI · p. 13

Abstract

from arXiv · show

Trust evaluation plays a critical role in cybersecurity by supporting risk mitigation and decision-making. A variety of trust evaluation methods have been proposed, with learning-based approaches offering high accuracy and automation. However, they often require substantial ground truth, suffer from low training efficiency, lack support for basic trust properties, and provide limited explainability. Large Language Models (LLMs) offer a compelling alternative due to their strong zero-/few-shot reasoning abilities and broad knowledge. To this end, we propose LLM4Trust, the first benchmark framework that systematically explores the capabilities of LLMs for trust evaluation. We first construct diverse trust graphs to model five basic trust properties and design corresponding property understanding tasks. We then assess the ability of eight representative LLMs to understand these properties under nine prompt methods. Based on this exploration, we identify the most effective LLM-prompt combinations and apply them to five real-world datasets for validating LLMs' trust evaluation capability. During this process, we propose two strategies to extract key information from large-scale trust graphs, addressing the context window limitations of LLMs. Extensive experiments show that LLMs can effectively understand basic trust properties and have great potential for real-world trust evaluation, particularly under limited supervision. However, they remain vulnerable to attacks targeting trust graphs and demonstration examples used in few-shot prompting, and incur high inference costs. Accordingly, we propose a defense mechanism and batch inference to improve the robustness and efficiency of LLM-based trust evaluation. The source code of LLM4Trust is available at https://github.com/Jieerbobo/LLM4Trust

I. INTRODUCTION

Learning-based trust evaluation can be accurate and automated, but often needs many labels, is computationally expensive to train, misses basic trust properties, and offers limited explanations. LLM4Trust benchmarks LLMs’ trust-property understanding and real-world evaluation, finding strong performance under limited supervision alongside robustness and cost challenges.

I. INTRODUCTION

  • Learning-based methods offer automated, high-performing evaluation but often need substantial ground-truth data, costly training, fail to capture trust properties, and provide little explanation.These limitations can make labels difficult to obtain in practice and undermine user acceptance.Source: §I · p. 1
  • Trust graphs face two evaluation challenges: existing datasets lack explicit property annotations, and LLM context windows restrict how much graph information prompts can include.The context limit may lead to incomplete reasoning or degraded performance in complex evaluations.Source: §I · p. 2
  • LLM4Trust evaluates five trust properties using synthetic graphs and question-answer tasks, testing eight LLMs with nine prompt methods.Three graph generators and varied graph parameters provide structural and task diversity.Source: §I · pp. 2–3
  • The study tests LLM trust evaluation on five real-world datasets against ten methods, using local subgraphs and recent-edge prioritization to fit relevant graph information into prompts.The two input strategies target context-window limits; recent-edge prioritization assumes newer behavior better reflects current trust.Source: §I · pp. 2–3
  • LLMs’ best-performing model-prompt combinations exceed 97% accuracy across all trust-property tasks and show advantages over existing methods under limited supervision.With sufficient labels, they remain less accurate than state-of-the-art learning-based methods; they outperform non-learning-based approaches.Source: §I · p. 2
  • LLMs can explain evaluations through step-by-step inference, but remain vulnerable to trust-graph attacks and poor few-shot demonstrations, while incurring high inference costs.A defense combining adversarial-example augmentation, temporal-edge filtering, and consistency correction mitigates the two attacks, recovering performance by up to 29.08%; batch inference reduces time per instance at some cost to accuracy and explainability.Source: §I · p. 2

A. Trust and Trust Evaluation

Trust evaluation predicts the trustworthiness of a trustor–trustee pair, while trust itself has multiple properties that shape how it should be evaluated. LLM4Trust addresses the lack of trust-specific LLM assessments through controlled property tasks and real-world evaluation.

A. Trust and Trust Evaluation

  • Trust evaluation predicts a trustor–trustee pair’s trustworthiness when its relationship is absent from the graph, and trust includes properties such as subjectivity, dynamicity, asymmetry, and composability.Trust evaluation quantifies trust using relevant factors.Source: A. Trust and Trust Evaluation · p. 3

B. Evaluation of LLMs

  • Existing LLM assessments leave trust-related concepts underexplored, while trust graphs’ imbalanced, multi-level, and temporal relationships make general graph benchmarks insufficient substitutes.LLM4Trust addresses this gap through knowledge-oriented and task-oriented assessments.Source: B. Evaluation of LLMs · pp. 3–4

A. Overview

  • LLM4Trust evaluates trust-property understanding (RQ1) and practical trust evaluation (RQ2), selecting LLM-prompt combinations before applying them with a tuned input format to real-world data.RQ1 compares responses against ground-truth answers; RQ2 extracts pair-associated subgraphs under hop-range and size constraints.Source: A. Overview · p. 4

B. Trust Graph Generators

  • Synthetic directed weighted trust graphs are generated with Erdős–Rényi by default, with Stochastic Block and Forest Fire models added for structural diversity.ER graph size is controlled by node count N, while edge probability p determines density; timestamps and trust levels are assigned for specific tasks.Source: B. Trust Graph Generators · p. 4

C. Property Understanding Task Generation

  • Five graph-based tasks operationalize dynamicity, asymmetry, conditional transitivity, composability, and context-awareness, with trust properties represented in controlled settings.Subjectivity is excluded because evaluations use objectively available data and interaction data already encode the trustor’s subjective perspective.Source: C. Property Understanding Task… · pp. 4–5
  • Task ground truths encode earliest trust-chain establishment, absent reverse trust, minimum trust along a chain, and no trust in a previously unseen context.Composability instead allows a trust range between plausible aggregation extremes because no single aggregation rule is universally accepted.Source: C. Property Understanding Task… · p. 5
  • Synthetic tasks provide controlled, literature-grounded property reasoning tests but do not fully capture practice’s complexities, motivating complementary evaluation on diverse, noisy real-world networks.Their reasoning patterns include temporal evolution, propagation, multi-source aggregation, and directional and context-dependent reasoning.Source: C. Property Understanding Task… · p. 5

D. Prompts

LLM4Trust varies prompting across demonstrations, external knowledge, and reasoning guidance, then applies selected models and prompts to real-world datasets. For context-limited graphs, it constructs pair-specific inputs using structural and temporal principles.

D. Prompts

  • Nine prompt types combine demonstration level, external knowledge, and reasoning guidance, including zero-, one-, and few-shot, knowledge, role, and chain-of-thought prompts.Few-shot uses three randomly selected examples by default; combined prompts pair CoT, knowledge, role, and few-shot components.Source: D. Prompts · pp. 5–6

E. LLMs

  • The evaluation uses seven general-purpose LLMs and GPT-5 for comparison, accessed through official APIs with temperature 0 and top p 1.0.Models were selected to vary in architecture, parameter size, context window, and knowledge cutoff.Source: E. LLMs · p. 6

F. Real-World Datasets

  • Five datasets cover social, certification, review, and Bitcoin networks, including static and timestamped dynamic graphs and binary or multi-level trust relationships.Bitcoin-OTC and Bitcoin-Alpha are dynamic; Advogato, PGP, and Epinions are static.Source: F. Real-World Datasets · p. 6

G. Best Input Format Construction

  • For each trustor–trustee pair, the input subgraph is constrained by hop range and edge count, with shorter-range neighbors prioritized for more direct trust information.The parameter study uses randomly sampled real-world data and the best-performing LLM-prompt combinations from RQ1.Source: G. Best Input Format Construction · p. 6
  • Dynamic-graph subgraphs prioritize recent interactions, and chronological splits distinguish nodes seen in both training and testing from nodes appearing only in testing.This temporal principle assumes recent interactions better reflect current trust states than outdated ones.Source: G. Best Input Format Construction · p. 7
  • When both pair members are observed, subgraphs follow the structural and temporal principles; with an unobserved trustee, a 2-hop out-degree subgraph captures the trustor’s preferences.The stated rationale for 2-hop neighbors is their sufficiency for trust propagation and aggregation in prior studies.Source: G. Best Input Format Construction · p. 7
  • With an unobserved trustor and observed trustee, a 2-hop in-degree subgraph represents the trustee’s objective trustworthiness; when both are unobserved, recent interactions proxy trust trends.The both-unobserved case is described as particularly challenging because neither party has trust-related data.Source: G. Best Input Format Construction · p. 7

IV. EVALUATION ON LLMS’ UNDERSTANDING OF BASIC TRUST PROPERTIES (RQ1)

LLMs can understand five basic trust properties, with performance shaped by prompt design, task difficulty, model capability, and graph complexity. Larger graphs and longer time spans are more challenging, while moderate numbers of examples and reasoning-oriented models can improve accuracy.

B. Main Results

  • Combined prompts, particularly few-shot+knowledge+role, outperform individual prompts; conditional transitivity and composability are harder than other properties.Few-shot examples illustrate graph-to-trust reasoning, while domain knowledge supplies explicit principles; propagation and aggregation rules contribute to task difficulty.Source: B. Main Results · pp. 7–8, Figure 3 · p. 8
  • Best LLM-prompt combinations achieve nearly 100% accuracy across trust-property tasks, with relative improvements of 27%–3133% over baselines.Some LLMs also perform well with zero-shot prompts.Source: B. Main Results · p. 7, Table III · p. 8
  • Claude-3.7-Sonnet reaches 100% accuracy on four tasks and 97% on composability with appropriate prompts.Larger, more recent models show advantages over smaller models with earlier knowledge cutoffs; GPT-5 performs better than general-purpose models but has substantially longer inference latency.Source: B. Main Results · p. 8, C. Impact of Trust Graph… · p. 9, Figure 4 · p. 9

C. Impact of Trust Graph Configurations, Shot Numbers, and Model Types

  • Conditional-transitivity accuracy stays above 90% across graph sizes and generator types, while composability performance degrades and varies more as graphs grow.Controlling input graph size is important for real-world use.Source: C. Impact of Trust Graph… · p. 9, Figure 5 · p. 9
  • Conditional-transitivity accuracy remains above 98% across trust-level and density settings, whereas composability is sensitive to these parameters and becomes harder in denser graphs with more trust levels.Dynamicity accuracy declines as time span increases, especially for large graphs.Source: C. Impact of Trust Graph… · p. 9, Figure 6 · p. 9, Figure 7 · p. 9
  • Accuracy rises with more demonstration examples on conditional transitivity and composability, but gains are limited or may reverse beyond three shots.Shot order has negligible effect; reasoning-oriented models improve accuracy, though GPT-5 incurs longer inference latency.Source: C. Impact of Trust Graph… · p. 9, Figure 8 · p. 9

V. EVALUATION ON LLMS’ CAPABILITY FOR TRUST EVALUATION (RQ2)

The real-world evaluation compares LLMs with established trust-evaluation methods on static datasets. LLMs perform especially well under limited supervision, while fully supervised learning methods remain stronger on some static datasets when ample labels are available.

A. Evaluation Setup

  • The evaluation uses F1-micro and MAE for static datasets, and F1-macro and balanced accuracy for dynamic datasets; it compares LLMs with representative statistical, inference, and learning-based baselines.The setup uses three LLMs with the few-shot+knowledge+role prompt and evaluates 1,000 sampled pairs per dataset across five random seeds.Source: A. Evaluation Setup · p. 10, §V · p. 10

B. Static Trust Graphs

  • Static input subgraphs balance retained information against input constraints: three-hop subgraphs work best for Advogato and PGP, while Epinions uses four hops.The respective edge limits are 60, 5, and 10; too few hops lose critical information, while too many may add noise.Source: B. Static Trust Graphs · p. 10, Figure 9 · p. 10
  • With three examples, LLMs achieve F1-micro scores of 0.673, 0.850, and 0.881 on Advogato, PGP, and Epinions, versus 0.518, 0.714, and 0.783 for the best 3-shot GNNs.LLMs also outperform non-learning methods and Matri across all three datasets.Source: B. Static Trust Graphs · pp. 10–11, Table IV · p. 11
  • With sufficient labels, LLMs remain less accurate than learning-based methods on Advogato and PGP, but reach state-of-the-art performance on Epinions.The passage notes that trust labels can require subjective ratings from thousands of entities, limiting the practicality of fully supervised methods.Source: B. Static Trust Graphs · p. 11
  • DeepSeek-V3 with three examples matches TrustGNN trained with about 1,500 labels on Advogato, 90,000 on PGP, and all 569,206 available labels on Epinions.DeepSeek-V3 without labels slightly outperforms its few-shot version; unrepresentative demonstrations may introduce variance.Source: B. Static Trust Graphs · p. 11, Figure 10 · p. 11

3) Explainability Analysis:

LLM-generated explanations improve participants’ trust-evaluation accuracy and confidence. The evaluation also finds that LLMs can perform competitively on dynamic datasets with few examples, but remain vulnerable to attacks on trust graphs and demonstrations; an input-level defense mitigates these effects.

3) Explainability Analysis:

  • LLM-generated explanations improve participants’ trust-evaluation accuracy by 28.97% and reported confidence by 24.73%.The user study involved 20 volunteers answering 12 questions with and without explanations.Source: 3) Explainability Analysis · p. 11

1) Input Format Study:

  • All three LLMs outperform dynamic methods under the same three-shot setting, while DeepSeek-V3 performs comparably to TrustGuard and better than other baselines.The dynamic datasets are difficult, particularly highly imbalanced Bitcoin-Alpha, and LLM performance varies substantially.Source: Table V · p. 12, 1) Input Format Study · p. 12
  • DeepSeek-V3 with three labeled examples matches TrustGuard trained with about 1,200 labels on Bitcoin-OTC and all 19,348 available labels on Bitcoin-Alpha.Without labels, DeepSeek-V3 is not competitive on either dynamic dataset.Source: 1) Input Format Study · p. 12, Figure 12 · p. 12

3) Robustness Analysis:

  • All LLMs suffer significant performance drops under mix, time-aware, and shot attacks, especially at high perturbation levels.On Bitcoin-Alpha, DeepSeek-V3’s absolute F1-macro drops are 10.79%, 22.30%, and 35.66% under the respective 30% mix, 30% time-aware, and 3-shot attacks.Source: Figure 13 · p. 12, 3) Robustness Analysis · p. 12
  • The input-level defense combines adversarial-example augmentation, temporal-edge filtering, and consistency correction, improving performance under all three attack types.It recovers performance by over 13% on average against time-aware attacks.Source: 3) Robustness Analysis · p. 13, Table VI · p. 13

VI. DISCUSSION AND LIMITATIONS

The discussion describes batch inference as a way to reduce LLM evaluation time, with trade-offs in accuracy and explainability, and illustrates LLM use during label-scarce Bitcoin trading periods. The conclusion highlights potential under limited supervision alongside vulnerability to attacks and high inference costs, while noting limits in expert guidance and graph scope.

VI. DISCUSSION AND LIMITATIONS

  • Batch processing reduces inference time per trustor-trustee pair, but large batches may slightly degrade accuracy and remove step-by-step explanations.LLMs typically take several seconds per pair, while learning-based approaches can evaluate thousands in under one second; total API cost for the main results was approximately $983.68.Source: §VI · p. 13, Table VII · p. 13
  • During Bitcoin-Alpha/OTC’s first three months, DeepSeek-V3 provides an initial evaluator with little task-specific supervision, while learning-based methods become more attractive as labels accumulate.For about 1,000 queried pairs, API costs were $0.40/$0.45 and average inference times were 10.24/5.66 seconds per query on Bitcoin-Alpha/OTC; LLM outputs may also bootstrap ML models as pseudo-labels.Source: §VI · p. 13
  • LLM4Trust’s expert guidance may be biased because trust rules are difficult to generalize across scenarios and individuals.The authors associate this limitation with Claude-3.7-Sonnet performing best on basic trust properties while DeepSeek-V3 performs better on real-world evaluation.Source: §VI · p. 13
  • Experiments evaluate trust from node interactions alone, leaving text-attributed trust graphs unevaluated because high-quality textual data are limited.The authors propose incorporating node descriptions into prompts as future work and expect this may improve accuracy.Source: §VI · pp. 13–14

VII. CONCLUSION

  • The conclusion reports strong potential for real-world evaluation under limited supervision, but also vulnerability to attacks on trust graphs and few-shot demonstrations and high inference costs.Future work will develop stronger defenses and improve inference efficiency for practical deployment.Source: §VII · p. 14

ETHICAL CONSIDERATIONS

  • The authors recommend restricting access to authorized stakeholders because LLM4Trust could be misused to infer trust relationships.The reported human-related studies received institutional review board approval, and participants gave informed consent.Source: ETHICAL CONSIDERATIONS · p. 14

APPENDIX A ALGORITHMS FOR GENERATING Q&A PAIRS

The appendices document procedures for generating question-answer pairs to assess five trust properties, alongside task-specific expert knowledge used in a prompt method.

APPENDIX A ALGORITHMS FOR GENERATING Q&A PAIRS

  • Algorithms 1–5 detail procedures for generating question-answer pairs for the five fundamental trust properties.The described properties include dynamicity, asymmetry, conditional transitivity, composability, and context-awareness.Source: Appendix A · p. 16, Appendix B · pp. 16–17

APPENDIX B DOMAIN EXPERT KNOWLEDGE

  • Table VIII presents task-specific domain expert knowledge used in the “knowledge” prompt.The context-awareness procedure assigns trust levels together with contexts to graph edges.Source: Appendix B · pp. 16–17, Table VIII · p. 17
  • The asymmetry procedure selects a one-way edge without a return path, while the conditional-transitivity procedure selects a reachable pair without a direct edge.The fragments also show that the composability procedure tracks minimum and maximum trust weights across paths.Source: Appendix B · p. 16

APPENDIX C PROMPT DETAILS

The appendix points readers to prompt templates and illustrates how a few-shot, knowledge, and role prompt is organized for an asymmetry-understanding task.

APPENDIX C PROMPT DETAILS

  • The example separates role, task, and operational instructions in the system prompt from task-specific input in the user prompt.Prompt templates for all methods are provided in Table IX.Source: Appendix C · p. 17

APPENDIX D ADDITIONAL EXPERIMENTS AND DETAILED SETUP

The appendix examines how demonstration-graph scale affects composability understanding and defines four metrics for evaluating trust predictions. Small-scale demonstrations generally retain high accuracy, while large-scale demonstrations reduce it.

A. Impact of Shot Scale

  • Accuracy is generally high with small-scale shots but declines noticeably with large-scale shots in the composability task.The authors suggest large shots may add noise that hinders learning effective evaluation rules; Table X reports accuracy under different shot scales.Source: A. Impact of Shot Scale · p. 17, Table X · p. 18

B. Evaluation Metrics

  • F1-micro aggregates class contributions into an overall F1 score, while F1-macro averages per-class F1 scores without weighting classes.F1-macro is less sensitive to class imbalance.Source: B. Evaluation Metrics · p. 17
  • Balanced accuracy averages true-positive and true-negative rates, while MAE measures the average absolute gap between predicted and ground-truth trust levels.The metrics assess both positive and negative classes equally for BA; MAE is defined over predicted and actual trust levels.Source: B. Evaluation Metrics · p. 17

C. Descriptions of Baseline Methods

This appendix describes static and dynamic trust-evaluation baselines, a user study of LLM explanations, and tests of the graph input format. The input-format experiments identify a shared optimal subgraph size for two dynamic datasets and report that structural and temporal design choices outperform ablations.

C. Descriptions of Baseline Methods

  • Static baselines range from trust propagation and subjective-logic inference to matrix factorization, neural methods, and GNNs.The described methods include MoleTrust, OpinionWalk, Matri, NeuralWalk, Guardian, GATrust, and TrustGNN.Source: C. Descriptions of Baseline Methods · p. 18
  • Dynamic baselines differ in temporal representation: TrustGuard and DTrust use discrete time, while Medley uses exact timestamps.TrustGuard uses attention, DTrust recurrent neural networks, and Medley attention; TrustGuard also includes a homophily-based attack defense.Source: C. Descriptions of Baseline Methods · p. 18

D. Experiments on Explainability

  • The explainability study compares participant accuracy and confidence on trust questions shown with or without LLM-generated explanations.Twenty volunteers answered questions from Advogato; explanations came from Claude-3.7-Sonnet, DeepSeek-V3, and Llama-4-Maverick.Source: D. Experiments on Explainability · pp. 18–19

E. Experiments on Input Format

  • Both dynamic datasets share an optimal subgraph size of 50 edges, and the full input-format design outperforms structural and temporal ablations.The ablations randomly select neighbor ranges for static data or omit time when constructing dynamic-data subgraphs; Tables XI and XII report the comparisons.Source: E. Experiments on Input Format · p. 19, Table XI · p. 19, Table XII · p. 19

F. Statistical Analysis

Statistical analyses find that LLM performance is more stable than few-shot GNN performance on static data and that DeepSeek-V3 significantly outperforms same-setting GNN baselines across the reported datasets. The appendix also examines robustness under attacks and temperature effects on output consistency and accuracy.

F. Statistical Analysis

  • DeepSeek-V3 significantly outperforms TrustGNN on Advogato (p = 0.0013), PGP (p = 0.0093), and Epinions (p = 0.0031), and TrustGuard on Bitcoin-OTC (p = 0.0012) and Bitcoin-Alpha (p = 0.0022).Under 3-shot settings, LLM standard deviations are 0.006 to 0.025 on static datasets and 0.023 to 0.061 on dynamic datasets; GNN variability reaches 0.169 on static data.Source: F. Statistical Analysis · pp. 19–20, Table XIII · p. 20, Table XIV · p. 20

G. Experiments on Robustness

  • Llama-4-Maverick generally shows higher attack robustness, while DeepSeek-V3 is vulnerable across most attacks.For DeepSeek-V3, 3-shot attacks substantially increase distrust predictions on Bitcoin-OTC, reflecting a conservative response when trust information is lacking.Source: G. Experiments on Robustness · p. 20, Figure 17 · p. 20, Figure 18 · p. 20
  • The proposed input-level defense combines adversarial-example augmentation, temporal-edge filtering, and consistency correction for different attack types.The designs target mix attacks, time-aware attacks, and shot attacks, respectively.Source: G. Experiments on Robustness · p. 20

H. Experiments on Temperature

  • Lower temperatures yield more consistent outputs, and Claude-3.7-Sonnet and DeepSeek-V3 show greater stability; accuracy has no clear correlation with temperature.The study measures inconsistency across repeated runs and reports that these two models also have the most stable accuracy across temperature settings.Source: H. Experiments on Temperature · p. 21

I. Experiments on Data Contamination

The appendix tests whether dataset exposure or node-ID memorization explains DeepSeek-V3’s performance, then discusses inference cost and comparisons with graph foundation models. Results suggest performance mainly reflects reasoning over trust graphs, while the authors also describe reasoning errors and limited hallucinations.

I. Experiments on Data Contamination

  • Dataset-name exposure and node-ID randomization produce no consistent performance increase or decrease across three static datasets, with results within the original 95% confidence intervals.LLMs recalled high-level dataset statistics but failed to recover specific trust relations; the authors conclude performance mainly comes from reasoning rather than memorizing individual relations.Source: §I · p. 21, Table XV · p. 21

J. Cost Analysis and Efficiency Comparison

  • LLMs avoid task-specific training and offer deployment flexibility, but typically need several seconds per trustor-trustee pair, unlike baselines processing thousands in under a second.The comparison excludes one-time pretraining costs and reports inference time for LLMs versus training overhead for learning-based approaches.Source: J. Cost Analysis and Efficiency… · p. 22, Table XVII · p. 22

APPENDIX E COMPARISON WITH GRAPH FOUNDATION MODELS

  • General graph models are not directly comparable because their task objectives, reasoning objectives, and evaluation settings differ from few-shot trust evaluation.The paper contrasts general graph tasks such as link prediction with directed, multi-level trust prediction and trust-property reasoning.Source: Appendix E · p. 22

APPENDIX F LLM HALLUCINATION

  • Observed hallucinations are limited, while most errors arise from incorrect reasoning over explicit trust graphs, such as misapplied propagation or aggregation.The authors suggest explicit graph inputs reduce ambiguity; they say reasoning errors may be alleviated with improved prompting and more capable models.Source: Appendix F · p. 22

Related papers

Loading 2609.33521v2…