Biological neural circuits are both modular and recurrent, with dense intra‑community connections and sparse inter‑community links. Most spiking neural networks (SNNs) used for reinforcement learning are feed‑forward, leaving open the question of how to obtain functional modules in a recurrent SNN and what benefits they bring. This work introduces Community Recurrent SNNs (CoRe SNNs): recurrent networks partitioned into communities with fixed roles. A controller community receives stimuli, each possible action has its own community, and latent communities store internal state. Spiking activity walks between communities, and the community reached determines the chosen action. Learning is local and three‑factor, making the recurrent communities resemble cortical microcircuits more closely than a feed‑forward actor. We compare three regimes for forming communities: dynamic modularity, where a new learning rule called Map‑STDP encourages modular structure during training; partial modularity, where a stochastic block model (SBM) initialization partially enforces it; and full modularity, where sealed communities enforce it completely. Experiments on CartPole with three spiking neuron models show that Map‑STDP improves learning and matches full modularity on two models, while also making the observed spike walk more compressible. Replacing the feed‑forward actor of a TD‑STDP actor‑critic with recurrent action communities improves performance on CartPole and LunarLander and matches it on Acrobot. The optimal degree of modularity depends on the task: full modularity helps most on Acrobot, whereas all three regimes perform similarly on LunarLander. Overall, these recurrent modular networks can match or exceed feed‑forward approaches, offering gains in bio‑plausibility and energy efficiency.
Review: The recurrent modular design bridges biological realism and reinforcement learning performance, delivering both accuracy and energy benefits.