A bidirectional Mamba model, referred to in the study as BM, recorded the lowest error among the tested methods when estimating the historical reliability of devices in a simulated wireless system. It also showed the smallest swings in those scores over time and the highest modeled value of task completion. The result comes from an NS-3 computer simulation with 500 devices, so it is evidence of comparative performance inside that setup, not a test of a live network.
The study's stated objective was to select a trusted collaborator that maximizes modeled value of task completion, or VoC. That makes VoC a modeled utility used for selection, rather than an independently observed task outcome. The paper tests whether BM ranks well on that measure as the minimum trust threshold changes.
A model built around behavior over time
BM combines several stages. It first builds a sequence of historical collaboration graphs across short time slots. A graph neural network, or GNN, then propagates and fuses device-reliability information across each network graph. The long-term stage uses a forward Mamba block and a backward Mamba block in each layer, integrating information across time slots. In plain language, the design combines local network evidence with a longer behavioral record.
The labels used for evaluation were generated inside the simulation. Performance data were recorded as tasks ran, and ground-truth reliability values were generated from recorded historical performance. BM was trained by minimizing cross-entropy, a loss measure based on the difference between computed reliability values and those ground-truth values.
Data were split into 80% training and 20% testing. Training used five-fold cross-validation and early stopping, while comparison results were averaged over 10 runs.
The simulated system executed 10,000 tasks sequentially, focusing on face recognition, with a default input size of 5 MB and a processing density of 2,339 cycles per bit. That gives the comparison a defined workload, but it also limits the reach of the result. A finding from this simulated task set does not by itself show how BM would perform on other tasks, network conditions or device populations.
The test's boundaries
BM was compared with LSTM, GNN and QS-Trust. It had the lowest root mean square error, or RMSE, and mean absolute error, or MAE, the two error measures used for the historical reliability assessment. The supplied account reports the ranking but not exact RMSE or MAE values, so the size of the difference cannot be determined from these results.
Across 500 time slots, BM showed the smallest fluctuation and consistently lower RMSE than the comparison algorithms. That is the study's clearest long-term finding: the model's reliability scores moved less while remaining below the other methods on the reported error measure. Because the test is simulated, however, the result does not establish that BM will track real-world behavior more reliably.
As the minimum trust threshold increased, VoC decreased for all four methods, while BM consistently had the highest VoC across the tested threshold levels. In the simulation, stricter trust requirements therefore coincided with lower modeled task-completion value for every approach. The pattern is consistent with the authors' interpretation that BM supports collaborator selection with higher modeled utility within this experiment, but it does not establish an effect on actual task completion.
The evidence also comes with a measurement caveat. The reported findings are qualitative trends: the supplied analysis does not give exact VoC values, confidence intervals or run-to-run dispersion. Averaging results over 10 runs provides a summary, but it does not reveal how much the outcomes varied from run to run.
The design leaves other questions open as well. The comparison covers three named baselines, and the supplied analysis describes no ablation separating graph fusion, bidirectional processing or Mamba. Without that component-by-component test, the results cannot show which part of the full architecture accounts for the differences. Nor is there an independent external test described.
The document is an arXiv version 1 preprint dated 25 Aug 2026. For researchers building trust and reputation systems for collaborative wireless or distributed computing, it is a focused simulation result: BM led the reported comparisons on accuracy, long-term stability and modeled VoC. Whether the ranking survives real-world traces, different workloads and independent measures of task outcomes remains unanswered.
Paper data and sources
Original title: Long-Term Behavioral Evaluation for Trusted Collaborator Selection via Bidirectional Mamba
Authors: Botao Zhu, Xianbin Wang
Journal/Repository: IEEE ICNC 2026
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text