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

TLDR

Language models improve both quantitatively and qualitatively with scale, yet their transformative capabilities remain poorly characterized. The study aims to clarify current and near‑future language model capabilities and limitations to guide research, anticipate disruptions, and mitigate harms. The authors created BIG‑bench, a 204‑task benchmark from 450 authors across 132 institutions covering diverse domains beyond current model limits, and evaluated GPT, dense, and sparse transformers of varying sizes against human baselines. Results show that scaling improves performance and calibration but still lags human baselines; models perform similarly across architectures with sparse models slightly better; gradual gains stem from knowledge or memorization, while breakthrough tasks require multi‑step reasoning; social bias rises with scale but can be reduced by prompting.

Abstract

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.