Zero-Shot Learning

Zero-shot learning studies how models can handle unseen classes, tasks, domains, or settings without labeled examples for the target. Traditional visual methods align image features with attributes or text embeddings, while newer approaches use vision-language models, large language models, prompt design, generated features, and test-time adaptation. Evaluation ranges from class-recognition benchmarks such as CUB and AWA2 to video localization, cross-domain transfer, and embodied tasks, often testing both unseen-class performance and retention of seen classes. A central tension is that broad transfer can be fragile: models may favor familiar classes, react strongly to prompt wording, or fail under distribution shift, while adaptation can improve results but weaken the claim of learning without target-specific training.