Everything Everywhere All at Once: LLMs can In-Context Learn Multiple Tasks in Superposition

TL;DR

This paper reveals that large language models can perform multiple distinct in-context learning tasks simultaneously during a single inference, demonstrating a superposition capability that expands understanding of their latent computational abilities.

Contribution

The study introduces and empirically validates the phenomenon of task superposition in LLMs, providing theoretical explanations and analyzing internal task composition mechanisms.

Findings

LLMs can perform multiple ICL tasks simultaneously in a single inference call.

Larger models can handle more tasks in superposition and calibrate outputs better.

Task superposition emerges even when models are trained on single tasks.

Abstract

Large Language Models (LLMs) have demonstrated remarkable in-context learning (ICL) capabilities. In this study, we explore a surprising phenomenon related to ICL: LLMs can perform multiple, computationally distinct ICL tasks simultaneously, during a single inference call, a capability we term "task superposition". We provide empirical evidence of this phenomenon across various LLM families and scales and show that this phenomenon emerges even if we train the model to in-context learn one task at a time. We offer theoretical explanations that this capability is well within the expressive power of transformers. We also explore how LLMs internally compose task vectors during superposition. Furthermore, we show that larger models can solve more ICL tasks in parallel, and better calibrate their output distribution. Our findings offer insights into the latent capabilities of LLMs, further substantiate the perspective of "LLMs as superposition of simulators", and raise questions about the mechanisms enabling simultaneous task execution.