Efficient Exploration at Scale

TL;DR

This paper introduces an online reinforcement learning algorithm that significantly enhances data efficiency in training large language models with human feedback, achieving over tenfold improvements with fewer labels.

Contribution

The authors present a novel online RLHF algorithm that uses reward modeling, epistemic neural networks, and information-directed exploration to drastically reduce data requirements.

Findings

Achieves over 10x data efficiency compared to offline RLHF.

Matches performance of 200K-label offline RLHF with fewer than 20K labels.

Extrapolated results suggest 1,000x efficiency gains at larger scales.

Abstract

We develop an online learning algorithm that dramatically improves the data efficiency of reinforcement learning from human feedback (RLHF). Our algorithm incrementally updates reward and language models as choice data is received. The reward model is fit to the choice data, while the language model is updated by a variation of reinforce, with reinforcement signals provided by the reward model. Several features enable the efficiency gains: a small affirmative nudge added to each reinforcement signal, an epistemic neural network that models reward uncertainty, and information-directed exploration. With Gemma large language models (LLMs), our algorithm matches the performance of offline RLHF trained on 200K labels using fewer than 20K labels, representing more than a 10x gain in data efficiency. Extrapolating from our results, we expect our algorithm trained on 1M labels to match offline RLHF trained on 1B labels. This represents a 1,000x gain. To our knowledge, these are the first results to demonstrate that such large improvements are possible.