Reproduce “Language Models Predict Empathy Gaps Between Social In-groups and Out-groups” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yu Hou, Hal Daumé III, Rachel Rudinger

Large language models consistently predict empathy gaps between social in-groups and out-groups, mirroring human psychological patterns and historical tensions across various identity categories. Models like Llama-3.1-8B showed robust in-group favoritism, assigning higher emotion intensity scores when perceiver and experiencer shared the same social identity.

Key insights

  • Large language models consistently exhibit intergroup empathy gaps, showing higher predicted emotion intensity for in-group members.
  • The design of system prompts, particularly those enforcing a specific persona, can amplify the observed intergroup empathy biases in LLMs.
  • LLMs form structured perceptions of inter-group relationships (e.g., nationalities), clustering countries based on cultural and geopolitical contexts.
  • The models' empathy predictions reflect real-world conflicts and historical animosities between certain national or religious groups.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.