From Detecting Agency to Doing Work: Self-Caused Credit Builds a Durable Behavioral Self in a Minimal Spiking Agent

The key insight isn't that spiking networks learn (they do), but that gating credit assignment through detected agency creates behavioral persistence independent of episodic memory. This suggests self-model formation and learning stability may be coupled mechanisms rather than separate problems.

This work sits in a different layer than recent coverage on optimizer dynamics and robustness. Where the Hessian eigenvector paper (late June) maps how loss landscapes shift during training, and the distributionally robust inverse problems work constrains adversarial noise to physics-aligned patterns, this paper asks a prior question: how does an agent decide what to learn from in the first place? The agency-gating mechanism is a credit assignment filter, closer in spirit to the anomaly detection work on embedding reliability, except applied to behavioral learning rather than latent space diagnostics.

If the same agency-detection mechanism generalizes to multi-agent or hierarchical settings (beyond the minimal single-agent setup), that confirms the approach scales beyond toy domains. Watch whether follow-up work from this group or independent teams applies the PES+agency framework to continuous control tasks on neuromorphic hardware within the next 12 months; if it doesn't, the contribution may be limited to proof-of-concept rather than practical embodied AI.