DAL-Chat-SSL: Deep active learning with LLM oracles and embedding-space entropy for multi-modal classification

Deep active learning reduces annotation costs by querying only the most informative samples, but it depends on human annotators who are often unavailable in data-scarce settings. We propose deep active learning with artificial intelligence chatbots for semi-supervised learning (DAL-Chat-SSL), a hybrid framework that replaces human annotators with large language model (LLM) chatbots as labeling oracles and extends this paradigm—previously confined to text classification—to tabular, time-series, and image modalities. The framework introduces an embedding-space entropy acquisition function based on inverse-distance weighting in the learned representation, designed to address the miscalibration of softmax outputs that limits conventional predictive entropy under small labeled pools. Two chatbots, ChatGPT and Claude, serve as annotation oracles under a documented prompting protocol; their annotation accuracy is measured on held-out probe sets rather than assumed. We integrate active acquisition with semi-supervised self-training and evaluate the framework on six benchmarks (Sonar, Ionosphere, Optdigits, ECG5000, MNIST binary, MNIST 10-class) over 10 independent runs each, using a held-out test split that never participates in training, acquisition, or model selection. Against a fully supervised reference trained on all available labels, the framework recovers 84.8–99.8% of attainable performance while annotating 4.4–37% of the pool with no human involvement; on Ionosphere, 46 chatbot annotations are statistically indistinguishable from full supervision on 208 ground truth labels. Annotation error rates of 14–33% cost between one and seven points of final performance, and per-class analysis confirms that no class is lost to annotation noise. Acquisition strategies are separate from random sampling on exactly one of the six benchmarks, the one on which the initial model is already close to its ceiling; we relate this to the position of the model on its learning curve rather than to the annotation budget alone. The results indicate that LLM oracles, combined with semi-supervised self-training, offer a practical route to label-efficient learning beyond the text domain.