A Taxonomy of Transcendence

TL;DR

This paper investigates how language models surpass individual human capabilities by analyzing training data properties, introducing a knowledge graph-based setting to study modes of transcendence like skill denoising, selection, and generalization.

Contribution

It presents a taxonomy of transcendence modes and a novel knowledge graph-based testbed for studying how data diversity enables model capabilities beyond source expertise.

Findings

Identified three modes of transcendence: skill denoising, selection, and generalization.

Developed a knowledge graph-based setting for controlled experiments.

Highlighted data diversity aspects that facilitate transcendent model capabilities.

Abstract

Although language models are trained to mimic humans, the resulting systems display capabilities beyond the scope of any one person. To understand this phenomenon, we use a controlled setting to identify properties of the training data that lead a model to transcend the performance of its data sources. We build on previous work to outline three modes of transcendence, which we call skill denoising, skill selection, and skill generalization. We then introduce a knowledge graph-based setting in which simulated experts generate data based on their individual expertise. We highlight several aspects of data diversity that help to enable the model's transcendent capabilities. Additionally, our data generation setting offers a controlled testbed that we hope is valuable for future research in the area.