Self-training uses a model's own predictions on unlabeled data as training labels for semi-supervised learning. Process: Train on labeled data → predict on unlabeled data → select high-confidence predictions → add as pseudo-labels → retrain on expanded dataset → iterate. Why it works: Model extracts patterns from unlabeled data structure, confident predictions often correct, bootstraps from small labeled set. Selection strategies: Confidence threshold, top-k predictions, curriculum (easy to hard), uncertainty sampling. Risks: Error propagation (wrong pseudo-labels reinforce errors), confirmation bias, domain shift between labeled/unlabeled. Mitigation: High confidence thresholds, noise-robust training, consistency regularization, multiple models. For NLP: Text classification, NER, sequence labeling, instruction tuning from raw text. Related methods: Co-training (multiple views), tri-training (multiple models), Mean Teacher. Noisy Student: Google's large-scale self-training for vision - student trained on noisy augmented pseudo-labeled data. Modern use: Distillation from large models, domain adaptation, low-resource scenarios. Foundational semi-supervised technique.