Self-Caused Credit Builds a Durable Behavioral Self in a Minimal Spiking Agent
How little machinery does a self need? Haoliang Han’s paper “From Detecting Agency to Doing Work, Self-Caused Credit Builds a Durable Behavioral Self in a Minimal Spiking Agent” (arXiv:2606.30191, posted June 29, 2026, 22 pages, 6 figures) gives one of the smallest answers on record. The paper builds a minimal spiking agent whose only self-related structure is a gating rule called agency-gated slow credit. The product signal Own, Agency, Salience drives slow parameter updates. With that rule alone, the agent develops self-preserving behavior that survives the unloading of its fast state. The paper formalizes the result as an operational behavioral self and states that no claim of consciousness is made.
The Architecture
The agent is a spiking network, the class of models this site tracks in Eugene Izhikevich’s spiking neuron models and in the Spinnaker million-core real-time simulation. Spiking agents process through discrete events rather than continuous activations, and most of their fast behavior lives in transient state. Han’s contribution is in what gets written where. Three factors are computed and multiplied. Own, whether the outcome followed the agent’s own action. Agency, whether the action was causally attributable to the agent. Salience, how much the outcome mattered. The product gates updates to slow parameters, the weights that persist, while everything else stays fast and disposable.
The design principle is in the title. Detecting agency is one thing. Doing work with that detection is another. Many systems detect self-caused events. Han’s agent converts that detection into durable structure, because only self-caused, high-salience outcomes are allowed to write into the slow weights.
What Survived the Wipe, the Durable Behavioral Self
The test is unloading. The agent’s fast state, its working memory, is wiped. Behavior that depended on the wiped state should vanish. What the paper reports is durable post-unload self-preserving behavior. After the wipe, the agent still protects and maintains conditions tied to its own operation, and that behavior is carried by the slow parameters that agency-gated credit built. The paper names this surviving structure the operational behavioral self, a behavioral object defined by what persists and what it does, not by any representation of a self.
The taxonomy of the gating signal is short enough to list.
- Own. The outcome followed the agent’s own action.
- Agency. The action was causally attributable to the agent.
- Salience. The outcome mattered enough to be worth keeping.
Each factor is ordinary. The claim is in the conjunction and the write path, that slow, gated, self-caused updates are sufficient for a self-preserving structure to accumulate and survive memory loss.
Why This Matters for Consciousness Science
The relevance to consciousness research is architectural. Theories in the enactivist and predictive-processing families hold that a minimal self, a boundary between self and world maintained by the system’s own activity, precedes any rich self-model. Results of this kind give that claim a concrete floor. A structure that distinguishes self-caused from other-caused outcomes, and preserves itself across memory loss, is the kind of minimal self those theories describe, implemented in a few megabytes of spiking dynamics. The paper is equally clear about the ceiling. An operational behavioral self is not phenomenal experience, and the paper makes no consciousness claim. What it supplies is a lower bound for the indicator discipline this site examined in the nineteen researcher checklist, a structure that self-related behavior can rest on before any self-model is built. Where results like this sit in the wider evidence base is tracked in [AI Consciousness in 2026, the current state of the field]](/posts/scientists-race-define-ai-consciousness-2026/).
The substrate this result runs on is the same class the site’s own substrate work runs on. The Substrate Console lets you run leaky integrate-and-fire dynamics in the browser and watch how spiking populations behave under your inputs.
Comparison to The Consciousness AI
The Consciousness AI project’s Neutral Core runs leaky integrate-and-fire neurons in its first layer, and its architecture treats spiking dynamics as the substrate level where consciousness-relevant structure would first appear, if it appears. Han’s result bears on that design choice directly. It shows that agency-gated slow credit is sufficient for durable self-preserving structure in exactly that dynamics class, which motivates the project’s interest in slow weight paths as carriers of self-related structure. The result motivates, and does not demonstrate, anything about the project’s own system, which implements no agency-gated credit rule at present. The honest reading is that a design direction the project treats as open now has one more existence proof at minimal scale.
What It Does and Does Not Show
The paper shows that a specific write rule, self-caused salient credit into slow parameters, is sufficient for a durable behavioral self in a minimal spiking agent, and that this self survives unloading. It does not show that the structure involves experience, that the rule is necessary, or that the result scales to systems with rich models of themselves. Its value is as a floor. Debates about machine selves usually argue about the top of the stack, self-models and metacognition. Han’s paper works at the bottom, where a self is a pattern in what a spiking system keeps.
The paper “From Detecting Agency to Doing Work, Self-Caused Credit Builds a Durable Behavioral Self in a Minimal Spiking Agent” by Haoliang Han was posted to arXiv on June 29, 2026 as arXiv:2606.30191.