EvalMORAAL: Interpretable Chain-of-Thought and LLM-as-Judge Evaluation for Moral Alignment in Large Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Mohammadi, Hadi, Giachanou, Anastasia, Bagheri, Robert A.
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910237808656384
author Mohammadi, Hadi
Giachanou, Anastasia
Bagheri, Robert A.
author_facet Mohammadi, Hadi
Giachanou, Anastasia
Bagheri, Robert A.
contents We present EvalMORAAL, a transparent chain-of-thought (CoT) framework that uses two scoring methods (log-probabilities and direct ratings) plus a model-as-judge peer review to evaluate moral alignment in 20 large language models. We assess models on the World Values Survey (55 countries, 19 topics) and the PEW Global Attitudes Survey (39 countries, 8 topics). With EvalMORAAL, top models align closely with survey responses (Pearson's $r \approx 0.90$ on WVS). Yet we find a clear regional difference: Western regions average $r=0.82$ while non-Western regions average $r=0.61$ (a 0.21 absolute gap), indicating a persistent regional alignment gap. Our framework adds three parts: (1) two scoring methods for all models to enable fair comparison, (2) a structured CoT protocol with self-consistency checks, and (3) a model-as-judge peer review that flags 348 conflicts using a data-driven threshold. Peer agreement relates to WVS survey alignment ($r=0.74$, $p<.001$; PEW $r=0.39$, n.s.), supporting automated quality checks. These results show real progress toward culture-aware AI while highlighting open challenges for use across regions.
format Preprint
id arxiv_https___arxiv_org_abs_2510_05942
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle EvalMORAAL: Interpretable Chain-of-Thought and LLM-as-Judge Evaluation for Moral Alignment in Large Language Models
Mohammadi, Hadi
Giachanou, Anastasia
Bagheri, Robert A.
Computation and Language
Artificial Intelligence
We present EvalMORAAL, a transparent chain-of-thought (CoT) framework that uses two scoring methods (log-probabilities and direct ratings) plus a model-as-judge peer review to evaluate moral alignment in 20 large language models. We assess models on the World Values Survey (55 countries, 19 topics) and the PEW Global Attitudes Survey (39 countries, 8 topics). With EvalMORAAL, top models align closely with survey responses (Pearson's $r \approx 0.90$ on WVS). Yet we find a clear regional difference: Western regions average $r=0.82$ while non-Western regions average $r=0.61$ (a 0.21 absolute gap), indicating a persistent regional alignment gap. Our framework adds three parts: (1) two scoring methods for all models to enable fair comparison, (2) a structured CoT protocol with self-consistency checks, and (3) a model-as-judge peer review that flags 348 conflicts using a data-driven threshold. Peer agreement relates to WVS survey alignment ($r=0.74$, $p<.001$; PEW $r=0.39$, n.s.), supporting automated quality checks. These results show real progress toward culture-aware AI while highlighting open challenges for use across regions.
title EvalMORAAL: Interpretable Chain-of-Thought and LLM-as-Judge Evaluation for Moral Alignment in Large Language Models
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2510.05942