A preprint reports that TrustFormer achieved 91.3% overall accuracy when choosing collaborators in a simulated collaborative-computing setting. The standard Transformer baseline reached 85.4%, while the LSTM baseline reached 80.7%, and TrustFormer was higher across all reported dimensions.
The top-line result came alongside lower trust-evaluation errors in comparisons involving simulated collaborators showing strategic behavior or gradual degradation. The authors describe the framework as a way to improve trust evaluation and collaborator selection within the tested setting.
A trust history built around each task
TrustFormer is a task-specific, multi-dimensional trust evaluation framework. It uses task identifiers and timestamps generated by devices to bring together heterogeneous historical data, then applies cross-temporal and cross-dimensional attention to model relationships across time and data dimensions.
The framework is designed for data that may be collected asynchronously. Its synchronization step uses task information and device-generated timestamps before the model evaluates the historical record.
The historical data combine collaborator-side monitoring of CPU and memory with task-owner records about validated completion accuracy and timeliness. That gives the framework both device-resource information and records of how previous tasks were completed.
The modeled reporting process assumes that collaborators and task owners report honestly. As a result, the reported setup does not test how the framework handles dishonest reporting.
The system also evaluates whether a device has the resources a task needs. Resource trust is binary: it is assigned 1 when reported capability meets the task requirement and 0 otherwise. Selection requires all trust thresholds to be met while maximizing aggregate trust.
From three device models to a virtual test
The researchers first profiled behavior using two collaborative tasks on three Dell device models: the DELL 5200, DELL 5820 and DELL 7060. They used five constructed behavior modes and collected 100 collaboration instances for each device, task and behavior combination.
Those profiles were expanded into a simulation containing 200 virtual devices and 20,000 records. The records were divided chronologically into 70% for training, 10% for validation and 20% for testing.
Training used a model width of 256, four attention heads and two block layers, with mean squared error as the training loss. The system was compared with an LSTM and a standard Transformer, using mean squared error and mean absolute error to measure trust-evaluation error.
The difference grew with harder behavior
For strategic collaborators, TrustFormer recorded a mean squared error of 0.0268. The paper reports that its evaluation error was 40.8% lower than the LSTM baseline and 28.3% lower than the standard Transformer baseline.
For collaborators showing gradual degradation, the paper reports mean squared error reductions of 37.5% and 22.1% compared with the two baselines. The reported method differences were small for reliable behavior but widened when collaborators displayed more complex behavioral dynamics.
The abstract separately reports a 40.8% improvement in trust-evaluation accuracy and more reliable collaborator selection. The detailed selection result was 91.3% for TrustFormer, compared with 85.4% for the standard Transformer and 80.7% for the LSTM.
Timing still mattered
The simulations also examined the effect of disorder in historical monitoring records. With synchronized data, mean squared error was 0.0181 plus or minus 0.0008, while mean absolute error was 0.0923 plus or minus 0.0021. At 50% disorder, those figures rose to 0.0301 plus or minus 0.0014 and 0.1213 plus or minus 0.0033.
The paper describes those changes as increases of 66.3% for mean squared error and 31.4% for mean absolute error. In the reported comparison, keeping the historical records synchronized was associated with lower evaluation error.
The reported figures come from profiles based on three physical device models, two tasks and five constructed behavior modes, expanded into a virtual-device simulation. They therefore describe performance within the tested dataset and modeled setting.
The authors conclude that TrustFormer performs better on trust evaluation and collaborator-selection reliability, especially when the simulated behavior involves strategic deception or gradual degradation. The study's honest-reporting assumption remains part of the conditions behind that conclusion.
The manuscript is a preprint labeled arXiv:2608.25238v1 and dated 25 August 2026.
Paper data and sources
Original title: TrustFormer: Cross-Temporal and Cross- Dimensional Transformer for Task-Specific Multi-Dimensional Trust Evaluation
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