Improving on out-of-the-box LLM performance

This paper demonstrates that co-training can improve the performance of prompt-based learning using unlabeled data. While prompting has emerged as a promising paradigm for few-shot and zero-shot learning, it is often brittle and requires much larger models compared to the standard supervised setup. Co-training makes it possible to improve the original prompt model while simultaneously learning a smaller, downstream task-specific model. In cases of partial access to a prompt model such as GPT-3, the method learns a calibration model over prompt outputs. With full access to gradients, it learns soft prompt vectors through iterative updates. Models trained this way significantly improve performance on challenging datasets where there is a large gap between prompt-based and fully supervised learning.