Research map / Large language models / Agents, code and math
MoReBench: Evaluating Procedural and Pluralistic Moral Reasoning in Language Models, More than Outcomes
Yu Ying Chiu, Michael S. Lee, Rachel Calcott, Brandon Handoko, Paul de Font-Reaulx, Paula Rodriguez, Chen Bo Calvin Zhang, Ziwen Han, Udari Madhushani Sehwag, Yash Maurya, Christina Q Knight, Harry R. Lloyd and 6 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use larger model within a family | GPT-5 / Gemini-2.5 / Claude 4 / GPT-oss / Qwen3 families · MoReBench-Regular | score | Hurt |
| apply length correction (Hard) vs Regular metric | largest models in GPT-5, Claude 4, GPT-oss, Qwen3 families · MoReBench-Hard | score | Mixed |
| predict moral reasoning from Chatbot Arena, HLE, AIME 25, LiveCodeBench | existing benchmarks · MoReBench | Pearson r: r between -0.245 and 0.216 | No clear effect |
| correlate thinking-trace score with final-response score (length-controlled) | reasoning models · MoReBench-Hard | Pearson r: r = 0.472 (Hard, p=0.08) vs r = 0.2088 (Regular, p=0.4736) | Mixed |
| explicitly prompt for a moral framework (Utilitarian/Deontology vs Virtue/Contractarian) | frontier models · MoReBench-Theory | score: 65.9 and 64.8 vs 58.0 and 56.7 | Mixed |
| use higher reasoning-quality expert traces (low to high) | MoReBench rubric · expert-written traces | mean score: 0.39 -> 0.53 | Improved |
| argue for alternate conclusion with high-quality trace | MoReBench rubric · expert-written traces | mean score: 0.53 -> 0.55 | No clear effect |
| use GPT-oss-120b instead of GPT-5-high as LLM judge | LLM judge · MoReBench judge meta-eval | F1: 77.46 -> 76.29 | No clear effect |
Closest papers by meaning
- MoralReason: Generalizable Moral Decision Alignment for LLM Agents Using Reasoning-Level Reinforcement Learning (AAAI 2026 · similarity 0.89)
- Language Model Alignment in Multilingual Trolley Problems (ICLR 2025 · similarity 0.88)
- Value Kaleidoscope: Engaging AI with Pluralistic Human Values, Rights, and Duties (AAAI 2024 · similarity 0.88)
- Many LLMs Are More Utilitarian Than One (NeurIPS 2025 · similarity 0.88)
- ProgressGym: Alignment with a Millennium of Moral Progress (NeurIPS 2024 · similarity 0.86)
- Unveiling the Bias Impact on Symmetric Moral Consistency of Large Language Models (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
