Reproduce “Value Kaleidoscope: Engaging AI with Pluralistic Human Values, Rights, and Duties” on OpenResearch

About this paper

Taylor Sorensen, Liwei Jiang, Jena Hwang, Sydney Levine, Valentina Pyatkin, Peter West, Nouha Dziri, Ximing Lu and 5 more

This research introduces VALUEPRISM, a large-scale dataset containing 218,000 pluralistic human values, rights, and duties derived from 31,000 real-life situations using GPT-4 and human validation. A multi-task language model called KALEIDO, trained on VALUEPRISM, generates and explains contextualized moral considerations, outperforming GPT-4 in human evaluations and demonstrating zero-shot transferability to philosophical ethical benchmarks.

Key insights

• Large language models (LLMs), even after distillation into smaller models, can effectively model and reason about complex ethical concepts, including pluralistic human values, rights, and duties, and their interactions.
• An interpretable, modular approach to value representation, like KALEIDO, allows for direct inspection of how values contribute to ethical judgments, enhancing accountability and understanding of AI's moral reasoning.
• The output entropy of KALEIDO's decision system can implicitly identify situations with high variability or ambiguity in human moral judgments, without explicit training for this feature.

Summary by alphaXiv.