Generating Compromises Between Two Points of View

ACL ARR 2025 February Submission5635 Authors

Abstract: Large Language Models (LLMs) excel academically but struggle with social intelligence tasks. We present a framework for generating empathicly neutral compromises between opposing viewpoints to address this limitation. Using a dataset of 2,400 contrasting views from human participants, we develope compromise generation methods. To overcome data collection constraints, we utilize prompt engineering with Claude LLM to generate compromises, validated through a 50-participant user study. Models trained on this dataset via preference alignment demonstrated effective compromise generation across multiple metrics. This work establishes a scalable approach to enhance LLMs' social intelligence through neutral conflict resolution.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: social LLM

Contribution Types: NLP engineering experiment, Approaches to low-resource settings

Languages Studied: english

Submission Number: 5635