This repository serves as a comprehensive, community-driven compilation of research papers focused on In-Context Learning (ICL) in large language models. It aims to provide researchers and practitioners with a structured overview of the field, covering key areas like model training, prompt tuning, analysis of influencing factors, working mechanisms, evaluation, and applications.

How It Works

The repository categorizes papers into logical sections, facilitating targeted exploration of ICL research. It covers foundational surveys, papers on pre-training strategies, prompt engineering techniques, analysis of ICL's behavior and influencing factors (e.g., pre-training data, inference stage choices), evaluation benchmarks, and emerging applications. The structure allows users to quickly identify relevant work based on their specific research interests within ICL.

Quick Start & Requirements

This is a curated list of research papers, not a software library. No installation or execution is required. All information is accessible via the README.

Highlighted Details

• Comprehensive categorization of ICL research, from foundational surveys to specific techniques and analyses.
• Includes links to papers (PDFs), project pages, and demos where available.
• Covers a wide range of ICL aspects including prompt tuning, chain-of-thought, and meta-learning.
• Features a dedicated section for challenges and future directions in ICL research.

Maintenance & Community

The project is actively maintained and community-driven, encouraging contributions for new papers and insights. A contributor list is maintained, acknowledging significant contributions.

Licensing & Compatibility

The repository itself is not software and thus not subject to software licensing. The content consists of links to research papers, which are governed by their respective publication licenses and copyright.

Limitations & Caveats

This is a curated list and may not be exhaustive. The "working in progress" nature means new papers are continually added, and existing categorizations might evolve. It does not provide code or tools for implementing ICL.