Self-Improving Intelligence: Learning From Unlabeled Data Across Domains

This review investigates self-improving intelligence, emphasizing mechanisms by which models enhance performance through unlabeled data. It synthesizes studies on self-training, self-instruction, and domain adaptation, with particular attention to audio large language models. Self-training frameworks iteratively label high-confidence examples, reducing dependence on supervised corpora. In speech processing, systems such as STAR and SI-SDA employ decoding stability and reinforcement-driven optimization to achieve quantifiable error reduction in source-free adaptation. In language modeling, self-instruct and self-play frameworks demonstrate that large models can generate synthetic supervision aligned with pretraining objectives. Comparative evidence from environmental modeling, recommendation, and circular economy research indicates similar dynamics of iterative pseudo-label refinement and cross-domain transferability. Across disciplines, results confirm that selfassessment, confidence-guided filtering, and synthetic data generation enable continuous adaptation without explicit supervision.