Vector-ICL: In-context Learning with Continuous Vector Representations

TL;DR

This paper introduces Vector-ICL, a method enabling large language models to perform in-context learning on continuous vector representations from diverse domains by aligning input vectors with the LLM's embedding space.

Contribution

The paper demonstrates that LLMs can effectively learn from continuous vectors across various tasks by using lightweight projectors, extending ICL beyond textual data.

Findings

Vector-ICL often outperforms few-shot ICL and domain-specific models.

Pretraining projectors with language modeling objectives is effective.

Vector-ICL works across multiple modalities and tasks.

Abstract

Large language models (LLMs) have shown remarkable in-context learning (ICL) capabilities on textual data. We explore whether these capabilities can be extended to continuous vectors from diverse domains, obtained from black-box pretrained encoders. By aligning input data with an LLM's embedding space through lightweight projectors, we observe that LLMs can effectively process and learn from these projected vectors, which we term Vector-ICL. In particular, we find that pretraining projectors with general language modeling objectives enables Vector-ICL, while task-specific finetuning further enhances performance. In our experiments across various tasks and modalities, including text reconstruction, numerical function regression, text classification, summarization, molecule captioning, time-series classification, graph classification, and fMRI decoding, Vector-ICL often surpasses both few-shot ICL and domain-specific model or tuning. We further conduct analyses and case studies, indicating the potential of LLMs to process vector representations beyond traditional token-based paradigms.