In a computer simulation, a model designed to help distributed systems choose collaborators scored better than the listed comparison methods on historical trust evaluation, according to a preprint. MVE recorded a Macro-F1 of 0.8247 ± 0.0058 and a mean absolute error, or MAE, of 0.0734 ± 0.0029 across five reported runs. Macro-F1 is a measure of how well categories are identified, while MAE describes the average size of prediction errors.
The work tackles a problem that becomes harder when trust judgments draw on a patchwork of past observations. The records may come from several task owners and may differ in source and quality. MVE treats each prior task owner as an independent observational view, puts that owner's records into a chronological time series, models patterns within each view with Mamba, estimates uncertainty with evidential deep learning, and dynamically fuses the resulting evidence.
Trust is one input to the decision
Historical behavior is not the only input to collaborator selection. MVE jointly considers a candidate's current resources and historical performance before recommending a collaborator. Its resource-trust component compares each collaborator-status dimension with the corresponding task requirement and uses a multiplicative calculation.
The test used a synthetic system
The evaluation used a simulated system with 200 devices and one server. It included three device types, five operational scenarios, 15 scenario-specific device models and two task types. The simulation included 200 task instances per device per scenario and 10,000 randomly executed collaborative tasks; the default task size was 500 MB.
Those five scenarios were mapped to ground-truth trust categories. Elite was paired with highly trusted at [0.9, 1.0), stable with trusted at [0.7, 0.9), strategic with non-stationary at [0.4, 0.7), selfish with constrained at [0.2, 0.4), and failed with malicious at [0.0, 0.2).
MVE led on the main measures
Historical trust was benchmarked with Macro-F1 and MAE. The comparison covered QS-Trust, TMC, DEF and MVE; DEF and TMC shared MVE's framework, with the neural-network backbone or fusion mechanism replaced.
Across five reported runs, QS-Trust had Macro-F1 of 0.5934 ± 0.0157 and MAE of 0.1635 ± 0.0079. TMC reached 0.7693 ± 0.0072 and 0.0912 ± 0.0038, while DEF reached 0.7865 ± 0.0068 and 0.0859 ± 0.0036. MVE led both measures, with 0.8247 ± 0.0058 for Macro-F1 and 0.0734 ± 0.0029 for MAE.
The authors reported that MVE outperformed DEF by 4.86% in Macro-F1 and 14.6% in MAE.
Uncertainty was part of the test
Uncertainty was part of the test, not just a feature of the model. In a retention analysis that ran from 100% of samples to 10%, MVE accuracy rose from 85.7% to 99.3% as higher-uncertainty cases were removed. DEF and TMC plateaued at 96.9% and 96.8%, respectively.
In a separate stress test, the paper introduced simulated conflict between views at inference, with the conflict ratio ranging from 0% to 50%. At 50%, MVE was reported to have degraded by 17.1% while retaining 68.6% accuracy. TMC degraded most severely, and DEF partly mitigated the decline.
Collaborator selection showed a similar directional result. As the plotted CPU-frequency and throughput requirements increased, MVE had a higher task-success rate than QS-Trust, and the gap widened progressively.
The result stays inside the simulation
The result has a narrow boundary. The benchmark covered two task types, three device types and five scenarios, with the trust categories tied to those scenario models. The evidence therefore describes performance under the reported simulation, not whether the same ranking would hold in a different system.
The manuscript is labeled arXiv:2608.25235v1 and dated 25 Aug 2026. Within that preprint, MVE led the listed methods on the reported trust measures and reached higher accuracy as the least-uncertain cases were retained. The work is best read as a simulated comparison of methods, with the results still to be tested beyond this setup.
Paper data and sources
Original title: Multi-View Trust Evaluation for Collaborator Selection via Evidential Deep Learning
Authors: Botao Zhu, Xianbin Wang
Journal/Repository: IEEE GLOBECOM 2026
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text