Transcendence: Generative Models Can Outperform The Experts That Train Them

TL;DR

This paper explores the phenomenon of transcendence, where generative models trained on human data surpass human experts in specific tasks, demonstrated through chess and supported by theoretical and experimental analysis.

Contribution

It introduces the concept of transcendence in generative models, providing theoretical proof and experimental evidence that low-temperature sampling can enable models to outperform their training data.

Findings

Transformers trained on chess data can outperform dataset players.

Low-temperature sampling facilitates transcendence.

Theoretical proof supports experimental results.

Abstract

Generative models are trained with the simple objective of imitating the conditional probability distribution induced by the data they are trained on. Therefore, when trained on data generated by humans, we may not expect the artificial model to outperform the humans on their original objectives. In this work, we study the phenomenon of transcendence: when a generative model achieves capabilities that surpass the abilities of the experts generating its data. We demonstrate transcendence by training an autoregressive transformer to play chess from game transcripts, and show that the trained model can sometimes achieve better performance than all players in the dataset. We theoretically prove that transcendence can be enabled by low-temperature sampling, and rigorously assess this claim experimentally. Finally, we discuss other sources of transcendence, laying the groundwork for future investigation of this phenomenon in a broader setting.