Emergence Conditions and Robustness Verification of Minimal Self-Awareness: A Systematic Study Based on Controlled Ablation Experiments

Abstract

Abstract The origin of self-awareness is a fundamental issue at the intersection of cognitive science and artificial intelligence. Existing studies mostly focus on the operating mechanisms of higher-order consciousness, and lack systematic controlled experimental verification of the emergence conditions for the most basal functional self. Targeting the core level of minimal functional self, this paper constructs the Eidolon‑S silicon-based Self-Interested Life Form simulation platform based on reinforcement learning, and designs multiple groups of single-variable controlled ablation experiments to systematically examine the functional hierarchy of three conditions — irreversible physical somatic boundary, multi-dimensional endogenous reward-punishment conflicts, and preconfigured multi-signal integration architecture — in the emergence of self-awareness. Aiming at the confounding variable of death semantics and training duration in boundary ablation experiments, a dedicated F0 Training Step Calibration Experiment is designed for conclusive verification. In addition, causal deprivation-reconstruction experiments are supplemented to examine its dynamic laws, and finally three rounds of methodological verification are conducted to rule out common methodological doubts. Conceptual definition: The minimal self-awareness in this paper specifically refers to the internal arbitration mechanism for coordinating multi-dimensional endogenous survival conflicts at the functional level, accompanied by somatic boundary sense and action control sense, corresponding to the minimal self level in cognitive science. It does not include higher-order constructs such as narrative self and phenomenal subjective experience. This paper takes the differentiation degree of attention arbitration as the operational observation indicator of minimal self-awareness. A total of 12 control groups are included in the study, all with a large sample of 2000 independent individuals. The baseline group verifies the emergence phenomenon; the B1 boundary ablation group, B2 boundary vulnerability ablation group, C conflict ablation group, and N homogeneous network ablation group respectively complete the single-variable test of the three core conditions; the F0 training step calibration group is dedicated to separating the confounding effects of death semantics and training duration; the N-Probe latent space probe group supplements the test of architectural induction bias; the E-series causal deprivation groups verify the sustaining role of controllability; the R-series gradient recovery groups examine the dose effect of injury-reconstruction; the S threshold scanning group, P hyperparameter robustness group, and M statistical correction group respectively respond to three types of methodological doubts: judgment threshold selection, hyperparameter tuning, and survivorship bias. The results show that: 1. Condition property stratification: Multi-dimensional endogenous reward-punishment conflicts and preconfigured multi-signal integration architecture are two core necessary conditions for the emergence of minimal self-awareness. The irreversible physical boundary acts as an evolutionary screening factor, not a necessary condition for emergence, and only serves to raise the emergence threshold. Weakened boundary vulnerability slightly reduces emergence intensity, but does not alter the essence of the phenomenon. 2. Confounding variable adjudication: After controlling for the training duration variable, the reversible death group still presents a 100% emergence rate and highly polarized attention weights, which directly rules out the alternative explanation of the training duration effect. It confirms that the boundary effect originates from the ontological semantics of death rather than differences in training experience, and verifies the existence of the pure fear constraint mechanism. 3. Dynamic laws: Causal controllability serves as both an emergence condition and a maintenance condition. Cutting off the causal chain triggers relative dissolution of the arbitration structure and learned helplessness-like survival collapse. Self-reconstruction capacity presents a negative dose-response relationship at the population level, and the irreversible effect is first manifested as survival elimination. 4. Methodological robustness: The baseline threshold lies at the midpoint of the phase transition interval, and the conclusions are free of post-hoc selection. The phenomenon is stably reproducible within a reasonable hyperparameter interval, and there exists a model capacity threshold. Arbitration function exerts a real mild positive effect on survival, and the observed strong correlation is mainly contributed by survivorship bias, with a bidirectional positive feedback between the two. This study forms a complete evidence chain of condition ablation – confounding calibration – dynamic test – methodological verification, providing reproducible computational experimental support for understanding the functional origin of minimal self-awareness. Keywords: Minimal Self-Awareness; Self-Arbitration; Consciousness Emergence; Multi-Objective Reinforcement Learning; Attention Mechanism; Ablation Experiment; Training Step Calibration