Scaling Laws for Pre-training Agents and World Models

TL;DR

This paper investigates how scaling up model size, data, and compute affects the performance of embodied agents, revealing power-law relationships similar to language models and highlighting factors influencing optimal model sizing.

Contribution

It characterizes the role of scale in world modeling and imitation learning, extending power-law insights from language modeling to these domains and analyzing influencing factors.

Findings

Power laws govern performance scaling in world modeling and imitation learning.

Tokenizer, task, and architecture significantly influence scaling coefficients.

Implications for optimal model and data sizing in embodied agents.

Abstract

The performance of embodied agents has been shown to improve by increasing model parameters, dataset size, and compute. This has been demonstrated in domains from robotics to video games, when generative learning objectives on offline datasets (pre-training) are used to model an agent's behavior (imitation learning) or their environment (world modeling). This paper characterizes the role of scale in these tasks more precisely. Going beyond the simple intuition that `bigger is better', we show that the same types of power laws found in language modeling also arise in world modeling and imitation learning (e.g. between loss and optimal model size). However, the coefficients of these laws are heavily influenced by the tokenizer, task \& architecture -- this has important implications on the optimal sizing of models and data.