True Few-Shot Learning with Language Models
This reveals that prior work overestimated the few-shot ability of language models, highlighting a critical challenge for researchers and practitioners in AI and NLP who rely on few-shot learning without access to validation data.
The paper tackles the problem of evaluating few-shot learning in language models without using held-out examples, finding that model selection criteria like cross-validation and minimum description length only marginally outperform random selection and often underperform significantly compared to using held-out examples.
Pretrained language models (LMs) perform well on many tasks even when learning from a few examples, but prior work uses many held-out examples to tune various aspects of learning, such as hyperparameters, training objectives, and natural language templates ("prompts"). Here, we evaluate the few-shot ability of LMs when such held-out examples are unavailable, a setting we call true few-shot learning. We test two model selection criteria, cross-validation and minimum description length, for choosing LM prompts and hyperparameters in the true few-shot setting. On average, both marginally outperform random selection and greatly underperform selection based on held-out examples. Moreover, selection criteria often prefer models that perform significantly worse than randomly-selected ones. We find similar results even when taking into account our uncertainty in a model's true performance during selection, as well as when varying the amount of computation and number of examples used for selection. Overall, our findings suggest that prior work significantly overestimated the true few-shot ability of LMs given the difficulty of few-shot model selection.