Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models

Language models demonstrate both quantitative improvement and new qualitativecapabilities with increasing scale. Despite their potentially transformativeimpact, these new capabilities are as yet poorly characterized. In order toinform future research, prepare for disruptive new model capabilities, andameliorate socially harmful effects, it is vital that we understand the presentand near-future capabilities and limitations of language models. To addressthis challenge, we introduce the Beyond the Imitation Game benchmark(BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450authors across 132 institutions. Task topics are diverse, drawing problems fromlinguistics, childhood development, math, common-sense reasoning, biology,physics, social bias, software development, and beyond. BIG-bench focuses ontasks that are believed to be beyond the capabilities of current languagemodels. We evaluate the behavior of OpenAI’s GPT models, Google-internal densetransformer architectures, and Switch-style sparse transformers on BIG-bench,across model sizes spanning millions to hundreds of billions of parameters. Inaddition, a team of human expert raters performed all tasks in order to providea strong baseline. Findings include: model performance and calibration bothimprove with scale, but are poor in absolute terms (and when compared withrater performance); performance is remarkably similar across model classes,though with benefits from sparsity; tasks that improve gradually andpredictably commonly involve a large knowledge or memorization component,whereas tasks that exhibit “breakthrough” behavior at a critical scale ofteninvolve multiple steps or components, or brittle metrics; social bias typicallyincreases with scale in settings with ambiguous context, but this can beimproved with prompting.

Further reading#

• Access Paper in arXiv.org