SPICA: Retrieving Scenarios for Pluralistic In-Context Alignment

Supplemental repository of code and data related to our paper SPICA: Retrieving Scenarios for Pluralistic In-Context Alignment.

Authors: Quan Ze Chen, K.J. Kevin Feng, Chan Young Park, Amy X. Zhang

Link: TBD

When aligning large language models (LLMs) to societal values, it is important to address a plurality of values reflected by diverse groups and communities. Existing in-context learning approaches for alignment often only consider similarity to the query when drawing few-shot examples, not accounting for cross-group differences around which values are prioritized.

In this work, we propose SPICA, a framework for pluralistic alignment that accounts for group-level differences during in-context example retrieval. We introduce three designs to facilitate pluralistic alignment: scenario banks, group-informed metrics, and in-context alignment prompts.

Overview

This repository contains

To cite our work, please refer to CITATION.cff or use the following:

@misc{chen2024spicaretrievingscenariospluralistic,
  title={SPICA: Retrieving Scenarios for Pluralistic In-Context Alignment},
  author={Quan Ze Chen and K. J. Kevin Feng and Chan Young Park and Amy X. Zhang},
  year={2024},
  eprint={2411.10912},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2411.10912},
}

Repository Layout

• ./annotation/: Contains frontend and backend code for human annotation.
• ./docs/: Contains code for the project website.
• ./data/: Contains non-identifying human annotations and evaluation data.