Learning Task Representations from In-Context Learning

TL;DR

This paper introduces a method to encode task information in in-context learning prompts by leveraging attention heads in transformers, improving task generalization across text and regression modalities.

Contribution

It presents an automated approach to derive task vectors from attention heads, enhancing understanding and generalization of task representations in large language models.

Findings

Effective extraction of task-specific information from demonstrations

Improved generalization across text and regression tasks

A new benchmark for evaluating task fidelity in ICL

Abstract

Large language models (LLMs) have demonstrated remarkable proficiency in in-context learning (ICL), where models adapt to new tasks through example-based prompts without requiring parameter updates. However, understanding how tasks are internally encoded and generalized remains a challenge. To address some of the empirical and technical gaps in the literature, we introduce an automated formulation for encoding task information in ICL prompts as a function of attention heads within the transformer architecture. This approach computes a single task vector as a weighted sum of attention heads, with the weights optimized causally via gradient descent. Our findings show that existing methods fail to generalize effectively to modalities beyond text. In response, we also design a benchmark to evaluate whether a task vector can preserve task fidelity in functional regression tasks. The proposed method successfully extracts task-specific information from in-context demonstrations and excels in both text and regression tasks, demonstrating its generalizability across modalities.