Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models
За пределами игры в имитацию: количественная оценка и экстраполяция возможностей языковых моделей
2022-06-09
SCID: 54.1/6nnyfxpw
Discuss with AI
BIG-bench benchmarklanguage model capabilitiesscaling lawssocial biassparse transformers
Figures from the paper
Abstract (AI)
Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.
Key Findings
1
Across models ranging from millions to hundreds of billions of parameters, performance and calibration generally improve with scale but remain poor relative to human expert raters.
2
BIG-bench introduces 204 diverse tasks contributed by 450 authors across 132 institutions to evaluate emerging and difficult language-model capabilities.
3
Gradual, predictable scaling is common for knowledge- or memorization-heavy tasks, while critical-scale breakthroughs often involve multistep reasoning, multiple components, or brittle metrics.
4
Performance is remarkably similar across dense and sparse model classes, although sparsity provides measurable benefits.
5
Social bias typically increases with scale in ambiguous-context settings, but prompting can reduce this effect.
Research Object
language models, including dense and sparse transformer architectures, evaluated across scales
Research Subject
their capabilities, limitations, scaling behavior, calibration, breakthrough performance, and social bias across diverse tasks
Publication Details
Publication Date
2022-06-09
Journal
Publisher
ISSN
Cited by
565
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest