Evan Ye

Independent Researcher, Shanghai, China

Functionalist What matters is the organisation of the processing, not the material

Evan Ye Photograph, courtesy of Evan Ye, used with permission.

Evan Ye works on the step where a system that only predicts the world starts to separate its own influence from everything else. His paper “From Prediction to Self, Developmental Conditions for Agency in Minimal Neural Systems” (arXiv:2606.05605, June 2026) traces that step in a 192 dimensional GRU across six experimental stages and 12 falsified alternatives. Its central finding is the encoding gap. A network can fully compensate for its own actions when it predicts, yet hold no readable state that says it is acting. Ye reports that the gap closes only when four conditions hold together, a persistent state with stable attractors, a causal loop from output back to input, proprioceptive feedback, and perceptual learning consolidated before action learning begins. He measures the result as agency gain, A = Err_world − Err_self, the drop in prediction error that comes from knowing one’s own action. After supervision is removed, the causally coupled network keeps its self encoding at 94.9% while a statistically matched control falls to 53.9%. Haoliang Han’s spiking agent, examined in the minimal spiking agent analysis, takes this detection step as its starting point.

The formal side of the program is set out in “The Epistemic State of an Observer, Negative Knowledge and Recovery from Wrong Internal Models” (Zenodo preprint, 2026). Ye defines an Observer Advantage, A_M, that scores an internal model against an observer who ignores the current state. A positive value is knowledge, zero is ignorance, and a negative value is what he calls negative knowledge, a model that predicts worse than having learned nothing. The paper proves that the capacity to know is bounded by log N while the capacity to err has no bound, and that under Bayesian updating the evidence needed to leave negative knowledge grows linearly with the strength of the wrong prior, n* = αc. Shannon mutual information falls out as the score of the ignorant observer. Agency gain is the share of A_M that comes from knowing one’s own action.

Ye places both papers inside a larger framework he calls Constraint Emergence Theory, introduced in his essay “The World You See Is Not the World Itself” (September 2026). Its claim is that an observer is a constrained subsystem of the world it models, so its representations differ from other observers’ while recovering the same underlying constraints. For consciousness science the work supplies a measurable floor under minimal-self accounts. Enactivist and predictive processing theories hold that a self and world boundary precedes any rich self-model, and Ye’s experiments specify the conditions under which a small network acquires an explicit version of that boundary. He makes no claim that such a network is conscious, and agency gain is defined on prediction error alone, so the measure applies to any substrate that predicts.

Known for. Agency gain, the encoding gap, developmental conditions for self-world separation in minimal neural systems, negative knowledge, Constraint Emergence Theory

Coverage on this site

Self-Caused Credit Builds a Durable Behavioral Self in a Minimal Spiking Agent September 2026

Related researchers

Evan Ye is listed with the other functionalists and computationalists, who hold that what matters is the organisation of the processing, not the material. The theory family behind that position is set out in the index of consciousness theories. Every researcher covered on this site is indexed in the directory of consciousness researchers.