New OMIT benchmark measures omission bias in LLM moral reasoning

Researchers have developed a new benchmark called OMIT to measure omission bias in large language models (LLMs) when they assist with moral reasoning. The benchmark includes 218 paired-frame scenarios across 10 conflict types, designed to capture disagreements among philosophical personas. Evaluations of eight LLMs revealed that omission bias is widespread but tends to decrease with larger model sizes within families. The study also found that interventions prompting models to consider moral principles before answering can reduce omission bias and improve frame consistency. AI

IMPACT This benchmark could lead to more robust LLM evaluations for moral reasoning, potentially improving AI safety and ethical decision-making.

RANK_REASON This is a research paper introducing a new benchmark for evaluating LLMs. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• CatalyzeX
• contractualism
• DagsHub
• deontology
• Gotit.pub
• Hugging Face
• Influence Flower
• LLMs
• ScienceCast
• utilitarianism
• virtue ethics

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →