SpikeMoE Copies Hippocampal CA1 Competition into a Spiking Mixture-of-Experts Router
Xiaoli Liu, Yujie Liang, Jialin Li, Chenxiao Dou, and Malu Zhang of the University of Electronic Science and Technology of China posted “SpikeMoE: Brain-Inspired Competitive Routing for Flexible Spiking Mixture-of-Experts” to arXiv on October 1, 2026 (arXiv:2610.01418, revised October 6). The paper takes the winner-selection circuitry of hippocampal region CA1, the k-winner-take-all competition implemented through lateral inhibition and refractory dynamics, and rebuilds it as the routing mechanism of a spiking mixture-of-experts model. The result selects Top-K experts by counting discrete spikes, and the reported benchmarks show state-of-the-art performance among spiking baselines, parity or better against dense artificial-network counterparts, and resilience when entire input modalities go missing.
The consciousness connection is direct and worth stating plainly. Selecting which experts fire is a selection-of-content problem, the same shape of problem as selecting which representations enter a global workspace. Mixture-of-experts routing is conditional computation, only some parts of the system are active for a given input. A spiking router selected by inhibitory competition makes that selection mechanism biologically grounded rather than arbitrary. The paper makes no claim that competitive routing supplies anything like conscious access, and that remains open. The substrate question this project tracks, which circuits can carry workspace-style selection on non-biological hardware, is exactly the question this architecture speaks to.
The mechanism, from CA1 to the router
Hippocampal CA1 solves a disambiguation problem. Many place cells respond to a given context, and recurrent inhibition prunes that set down to a stable, sparse winner set. The paper copies three elements of that circuitry into the router. Lateral inhibition, where active units suppress their neighbors. A refractory period, where a unit that just fired cannot immediately fire again, which prevents one expert from monopolizing every batch. And discrete spike counting, where routing decisions read integer spike counts rather than continuous activations, keeping the router itself inside the spiking regime.
The full stack then combines neuronal-scale spiking dynamics with model-scale expert selection. On top of the router the authors add a two-stage missing-modality module, which handles incomplete multisensory inputs by combining empirical prototypes drawn from an observed-modality pool with modality-specific learnable embeddings to synthesize representations for the missing modality. Vision, language, and multimodal benchmarks all test the framework.
| Design element | Biological origin | Engineering role in SpikeMoE |
|---|---|---|
| k-winner-take-all | Hippocampal CA1 competitive inhibition | Top-K expert selection |
| Lateral inhibition | CA1 interneuron circuits | Suppresses rival experts during routing |
| Refractory period | Neuronal firing reset | Prevents expert monopolization across batches |
| Spike-count routing | Discrete neural codes | Keeps the router in the spiking regime |
| Missing-modality module | Multisensory cortical integration | Synthesizes absent-modality representations |
Where this sits in the spiking substrate debate
Spiking neural networks are the closest machine analog to the temporal dynamics biological consciousness theories build on. The Substrate Console runs leaky integrate-and-fire neurons in Layer 1 precisely because LIF is the baseline event-driven unit that neuromorphic hardware implements, and the console exists to test what such dynamics can self-organize. SpikeMoE extends that landscape upward one level, from single-neuron dynamics to architecture-scale routing, showing that the competitive-inhibition motif scales to a modern sparse-computation workload.
The energy story is the reason the paper is more than an architecture note. Event-driven computation fires only when input arrives, which is the property that makes spiking hardware candidates for embedded and orbital deployment, the constraint profile examined in the neuromorphic satellite intelligence review. A routing layer that spends spikes to select experts, rather than dense matrix multiplies to evaluate them all, compounds that advantage. The paper reports a favorable trade between performance and energy efficiency as its headline result.
The hippocampal grounding also connects to the self-modeling line of work this site tracks. Han’s minimal spiking agent with a durable behavioral self built selfhood from slow self-caused credit updates in an LIF network, and the spiking default-mode-network pacemaker model tied elementary selfhood to tonic drive. SpikeMoE adds a different piece, a circuit motif for selective routing rather than for persistence or self-reference. The three together sketch the shape of a spiking cognitive architecture, dynamics, self-maintenance, and selection, built from separately validated pieces.
What the paper does not settle
The paper claims energy efficiency and benchmark performance, and it claims biological inspiration for the router. It does not claim that spike-count routing is how the hippocampus selects content for consciousness, nor that any of its benchmarks measure a consciousness-relevant property. The k-WTA analogy is at the circuit-mechanism level, CA1 does support k-winner-take-all dynamics, but the mapping from that motif to workspace access is a theory-laden step no benchmark in the paper takes. And the missing-modality module is engineered for robustness, not modeled on any biological multisensory-recovery process.
Those are the right limits to read the result inside. What the paper establishes is that a circuit motif long discussed in consciousness science runs efficiently as a routing mechanism at model scale, which makes the motif testable in artificial substrates at a scale the biology alone cannot reach.
Selection as the testable seam
If a global workspace needs a selection mechanism, SpikeMoE supplies a candidate implementation whose cost profile is measured. The testable question for the substrate program is whether workspace-style selection can be distinguished from plain load balancing in such architectures, by measuring what the selected set does downstream that an energy-equivalent random or dense router does not. The substrate workbench this project maintains exists to run exactly that kind of comparison on the LIF baseline, and the broader field context for why substrate-level selection evidence is traceable in the flagship overview of AI consciousness research. The open question is whether spike-count competition carries content selection or only traffic control, and that is a measurement, not an architectural claim.