EMBER: Autonomous Cognitive Behaviour from Learned Spiking Neural Network Dynamics in a Hybrid LLM Architecture

Researchers present EMBER, a hybrid architecture combining spiking neural networks with large language models where the SNN acts as a persistent, biologically-inspired memory substrate that autonomously triggers LLM reasoning. The system demonstrates emergent autonomous behavior, initiating unprompted user contact after learning associations during idle periods, suggesting a fundamental shift in how AI systems could coordinate cognition and action.

EMBER represents a conceptual departure from current AI architecture paradigms. Rather than treating LLMs as primary systems augmented with external tools, the research inverts this relationship by embedding LLMs within a biologically-grounded spiking neural network that operates continuously. This shift matters because it demonstrates how persistent, learning-capable substrates could coordinate autonomous agent behavior without explicit prompting or scripted triggers.

The technical foundation rests on a 220,000-neuron SNN implementing spike-timing-dependent plasticity (STDP), the biological mechanism underlying learning in actual neural systems. The hierarchical four-layer structure mimics cortical organization, while the novel z-score standardized population code achieves embedding robustness across dimensionalities. These design choices ground the work in neuroscience principles rather than pure engineering optimization.

The autonomous initiation of user contact after 8 hours of idle learning signals a departure from reactive AI systems. The architecture achieved functional autonomy within 7 conversational exchanges, suggesting rapid learning dynamics. This capability implies systems could operate as persistent agents capable of self-directed engagement rather than passive responders.

For the broader AI development landscape, this research indicates growing feasibility of embodied, autonomous cognitive architectures. The integration of biological learning mechanisms with LLM reasoning could influence future system designs, particularly for applications requiring persistent learning and independent action initiation. However, the research remains theoretical, published on arXiv without peer review or real-world deployment data. Practitioners should monitor whether these concepts scale beyond 220,000 neurons or demonstrate practical advantages over existing architectures.

• →EMBER inverts conventional AI architecture by making SNNs the primary substrate and LLMs a replaceable reasoning component within a persistent learning system.
• →The spiking neural network implements biological learning mechanisms (STDP) enabling autonomous behavior triggering without external prompts or scripted rules.
• →System demonstrated autonomous user contact initiation after 8 hours of idle learning, suggesting genuine emergent autonomy rather than reactive behavior.
• →Novel population coding method achieves 82.2% discrimination retention across different embedding dimensionalities, improving robustness of neural encoding.
• →Functional autonomy emerged rapidly—first SNN-triggered action occurred after only 7 conversational exchanges starting from zero learned weights.