View PDF HTML (experimental)

Abstract:How does a system that merely predicts the world come to distinguish its own causal influence from everything else? We trace this transition in a minimal 192-dimensional GRU through a developmental sequence -- 6 experimental stages, 12 falsified alternatives, and cross-signal validation. Starting with no action or self-representation, we add components one at a time, tracking whether the system distinguishes self-caused from world-caused changes. The central finding is the encoding gap: a system can perfectly compensate for its own actions in prediction while failing to encode "I am acting" as a readable state -- implicit causal use and explicit self-representation are dissociated capabilities.
The developmental path crosses this gap when four conditions are jointly satisfied: (1) persistent state that forms stable attractors, (2) a causal action loop linking the system's output to its input, (3) proprioceptive feedback that makes implicit causal knowledge explicit, and (4) asynchronous awakening -- consolidating perceptual learning before action learning, which yields the only configuration robust to hyperparameter choice. We propose agency gain (A = Err_world - Err_self), the predictive advantage of knowing one's own action, as a continuous metric that generalizes across signal types.
A decisive test confirms the causal grounding of the encoding: after the external training signal is removed, the causal agent retains its self-representation at 94.9% while a statistically-matched control collapses to 53.9%. Self-representation persists only when causally useful for prediction -- an intrinsic property of the causal loop, not a training artifact.

Submission history

From: Rui Ye [view email]
[v1] Thu, 4 Jun 2026 02:27:43 UTC (333 KB)
[v2] Thu, 20 Aug 2026 13:48:44 UTC (356 KB)