Lu et al. (2022), "Fantastically Ordered Prompts and Where to Find Them," study how the order of few-shot examples within an otherwise-identical prompt affects accuracy. What did they find, and what method did they propose to pick a good order without a labeled validation set?
- Example order has no measurable effect on accuracy once the examples themselves are held constant, so the paper concludes ordering can safely be ignored
- The order of the same few-shot examples can swing accuracy from near state-of-the-art to close to random guessing; they proposed generating an artificial, unlabeled "probing" set from the language model itself and selecting the ordering whose predicted-label distribution has favorable entropy statistics on that set, without needing any labeled dev data
- They found the best fix was to always sort examples alphabetically by their label text, which fully eliminated order sensitivity across all models and tasks tested
- They found order sensitivity only affects models under one billion parameters, and it disappears automatically once a model is scaled past that size
Why B? And why not the others?
Correct answer: B. The order of the same few-shot examples can swing accuracy from near state-of-the-art to close to random guessing; they proposed generating an artificial, unlabeled "probing" set from the language model itself and selecting the ordering whose predicted-label distribution has favorable entropy statistics on that set, without needing any labeled dev data
Lu et al. (2022) demonstrate that, holding the same few-shot examples constant and only permuting their order, GPT-family models can swing between near state-of-the-art and near-random accuracy, and that this sensitivity is unpredictable from prompt length or example choice alone. To pick a good ordering without labeled validation data, they use the generative model itself to construct an artificial probing set of unlabeled outputs and then select the candidate ordering whose predicted-label distribution over that probing set has favorable entropy statistics, reporting a 13% relative improvement on average across eleven text classification tasks. The second option is wrong because the paper's entire premise and headline finding is that order sensitivity is large and highly consequential, not negligible. The third option is wrong because no universal alphabetical-sort rule is proposed or shown to eliminate the effect; the paper's actual proposal is the entropy-based probing method. The fourth option is wrong because the paper does not report or claim that scaling past one billion parameters resolves order sensitivity.
Source: Lu, Bartolo, Moore, Riedel, Stenetorp, "Fantastically Ordered Prompts and Where to Find Them: Overcoming Few-Shot Prompt Order Sensitivity" (ACL 2022), arXiv:2104.08786