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

**Authors:** Yufan Zhuang, Chandan Singh, Liyuan Liu, Jingbo Shang, Jianfeng Gao  
**arXiv:** [2410.05629](https://arxiv.org/abs/2410.05629)  
**Date:** 2024-10-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2410.05629v1/

## Key points

- Large Language Models (LLMs) can extend in-context learning (ICL) capabilities to continuous vectors from diverse domains
- Vector-ICL aligns input data with LLM's embedding space through lightweight projectors for effective processing and learning
- Pretraining projectors with general language modeling objectives facilitates Vector-ICL, while task-specific finetuning enhances performance
- Vector-ICL consistently outperforms few-shot ICL and domain-specific models or tuning in experiments and case studies
- LLMs show potential in text summarization and molecule captioning tasks with evaluation metrics like RougeL and BLEU score
- LLMs are investigated for text reconstruction and arithmetic/function regression tasks using synthetic datasets
- Research highlights the versatility of LLMs in processing diverse data types through Vector-ICL methodology for applications beyond traditional token-based paradigms

## Summary

This study explores the potential of Large Language Models (LLMs) to extend their remarkable in-context learning (ICL) capabilities beyond textual data to continuous vectors from diverse domains. The researchers introduce Vector-ICL, a method that aligns input data with an LLM's embedding space through lightweight projectors, enabling effective processing and learning from these projected vectors. They find that pretraining projectors with general language modeling objectives facilitates Vector-ICL and task-specific finetuning further enhances performance. Through various experiments and case studies, Vector-ICL consistently outperforms few-shot ICL and domain-specific models or tuning. Additionally, the study showcases LLMs' potential in text summarization and molecule captioning tasks using performance evaluation metrics like RougeL and BLEU score. The researchers also investigate LLMs' ability in text reconstruction and arithmetic/function regression tasks using synthetic datasets. Overall, this comprehensive research highlights the versatility of LLMs in processing diverse data types through Vector-ICL methodology and their potential in various applications beyond traditional token-based paradigms.