A proposed spike-based controller reached the goal in the Mountain Car benchmark and then overshot it, as did the conventional numerical active-inference controller used for comparison. The authors describe the outcome as successful task completion, but the reported comparison was qualitative rather than a numerical performance comparison.
The document is an arXiv preprint, version 1, dated 20 Aug 2026. Its central question was whether a spike-based realization of a nonlinear state-transition factor could be integrated into belief propagation without degrading the original active-inference controller's behavior.
Inside the computational test
The reported validation focused on the Mountain Car problem, a classical benchmark in control theory and reinforcement learning. It compared the proposed spike-based implementation with a reference active-inference controller that relied on conventional numerical computations. This was a computational controller comparison, not a human or animal study.
The modeled state consisted of the car's position and velocity. To update beliefs about that state, the system used local sum-product message passing on a factor graph, breaking the calculation into linked local updates.
Continuous belief parameters were represented through population coding, using the joint activity of neural populations. Nengo implemented the spike-based nonlinear functions and their encoding and decoding, while RxInfer handled message passing and Unscented Kalman Filter approximations for the nonlinear messages.
The experiment reported the parameter values u0 = 0, y* = 0.4, Ng = 300, Nf = 100, Na = 100 and Flim = 0.04.
Similar behavior, with an important caveat
The comparison tracked vehicle position and engine-control behavior. Both systems initially moved away from the goal to build momentum before ascending the hill, giving the proposed controller the same broad action pattern as the reference.
Both methods then reached and overshot the goal position, which the paper treated as successful completion of the task. The supplied analysis does not give a numerical measure of how far they overshot.
The abstract also reports successful real-time state updates and goal-directed action plans generated through spike-driven dynamics. That claim concerns model behavior in the benchmark; it is not a report of deployment on neuromorphic hardware.
What the test leaves open
The authors interpret the findings as indicating that spike-based neural dynamics can support nonlinear Bayesian inference and control in a biologically inspired manner. The study does not test whether that approach outperforms the reference controller, because its comparison examines qualitative position and engine-action behavior rather than seeking exact trajectory reproduction.
The reported implementation is not fully spike-based: fully spike-based factor nodes and message passing were identified as future work. The paper also reports no energy-efficiency or neuromorphic-hardware evaluation.
The evidence is limited to the Mountain Car benchmark, and the state-transition model was hand-crafted rather than learned from environmental interactions. Repeated quantitative comparisons with additional controllers, more complex nonlinear environments and learned dynamics models would be needed to assess robustness and generalization.
Funding was acknowledged from the Eindhoven Artificial Intelligence Systems Institute at Eindhoven University of Technology.
Paper data and sources
Original title: Spike-based Belief Propagation in Nonlinear Dynamical Systems
Authors: Sepideh Adamiat, Hongye Wang, Wouter M. Kouw, Bert de Vries
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text