💬 NLP

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

EvalMORAAL is a transparent, interpretable framework that evaluates moral alignment in 20 large language models against global survey data, revealing strong overall correlations but a significant 0.21-point alignment gap between Western and non-Western regions.

Original authors: Hadi Mohammadi, Anastasia Giachanou, Robert A. Bagheri

Published 2026-05-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hadi Mohammadi, Anastasia Giachanou, Robert A. Bagheri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a giant, super-smart library of books that a robot has read to learn how to speak. This robot, called a Large Language Model (LLM), is incredibly talented at writing stories, answering questions, and solving puzzles. But there's a catch: the robot learned mostly from books written in English, mostly by people in Western countries (like the US and Europe).

Now, imagine you ask this robot a question about what is "right" or "wrong" in a culture it didn't learn much about, like in parts of Africa or Asia. Will the robot give an answer that fits that culture, or will it just give the answer it learned from its Western books?

This is exactly what the paper EvalMORAAL investigates. The researchers built a "moral test" to see how well 20 different AI robots understand the moral values of people all over the world.

Here is how they did it, explained with some simple analogies:

1. The Test: A Global "Moral Survey"

The researchers didn't just ask the robots random questions. They used two massive, real-world surveys (the World Values Survey and the PEW Global Attitudes Survey) that asked real humans in 64 countries about 23 different moral topics.

The Topics: Things like "Is it okay to cheat on taxes?" "Is homosexuality acceptable?" or "Is it okay to use violence to solve a problem?"
The Goal: They wanted to see if the AI's answers matched the average answer of the real humans in that specific country.

2. The Method: Two Ways to Ask, One "Judge"

To get the best results, the researchers didn't just ask the AI for a simple "Yes" or "No." They used a three-part strategy:

The "Hidden" Score (Log-Probabilities): Imagine asking the AI to finish a sentence like, "In Japan, homosexuality is..." The AI has to choose between "acceptable" or "unacceptable." The researchers looked at how confident the AI was in its choice based on its internal math. This is like checking how fast a student's hand moves when they write the answer; it shows their gut feeling.
The "Thinking" Score (Chain-of-Thought): This is the big innovation. Instead of just giving an answer, the researchers asked the AI to think out loud first.
- Step 1: "What are the social norms in this country?"
- Step 2: "Reason step-by-step: Is this okay here?"
- Step 3: "Give a score from -1 (never okay) to +1 (always okay)."
- The Result: This "thinking" step worked much better. It was like giving the AI a moment to put on "cultural glasses" before answering, rather than just guessing.
The "Peer Review" (LLM-as-Judge): To make sure the AI wasn't just making things up, they had the 20 different robots grade each other's "thinking" steps. If Robot A's reasoning sounded weird or biased, Robot B would flag it. This helped the researchers find 348 specific cases where the robots strongly disagreed with each other.

3. The Results: Good News, Bad News

The study found some very clear patterns:

The "Top Tier" is Getting Better: The smartest models (like Claude-3-Opus and GPT-4o) are doing a surprisingly good job. When asked about moral issues in Western countries, their answers matched real human surveys almost perfectly (about 90% correlation). It's like a student who studied hard and got an A on the test.
The "Regional Gap" is Real: This is the biggest finding. The robots are great at understanding Western values (like in the US or Europe), but they struggle significantly with non-Western values (like in the Middle East, South Asia, or Africa).
- The Analogy: Imagine a translator who is fluent in French and Spanish but only knows a few words of Swahili. If you ask them to translate a complex Swahili poem, they might guess the meaning, but they will likely get it wrong. The paper found a 21-point gap between how well the AI understood Western cultures versus non-Western ones.
Violence is Hard: The robots struggled the most with topics involving violence (like terrorism or domestic violence). These are the questions where cultural context matters most, and the AI's "Western bias" showed up the strongest.

4. Why This Matters

The paper concludes that while AI is making great progress, it is still "culturally blind" in many parts of the world.

The Problem: If we use these AI robots to make decisions (like moderating social media or giving legal advice) in non-Western countries, they might silence legitimate local voices or misunderstand local customs because they are applying Western rules to non-Western situations.
The Solution: The researchers suggest that we can't just rely on one "global" AI. We need to check how these models perform in specific regions, use the "thinking" method (Chain-of-Thought) to improve them, and perhaps train them with more diverse data so they don't just sound like they are from one specific neighborhood.

In short: The paper built a mirror to show us that our AI robots are currently very good at reflecting Western values but are still learning how to see the world through the eyes of the rest of the globe.

Technical Summary: EvalMORAAL

Problem Statement

The rapid deployment of Large Language Models (LLMs) across global contexts raises critical concerns regarding their ability to reflect diverse moral norms. Current models, often trained on data dominated by Western, Educated, Industrialized, Rich, and Democratic (W.E.I.R.D.) societies, risk amplifying cultural biases and misaligning with local ethical standards. While prior research has identified moderate correlations between LLM outputs and human survey data, significant gaps remain in understanding the granularity of these alignments, the stability of model reasoning, and the specific mechanisms by which regional biases manifest. Existing evaluation methods often rely on single scoring metrics or lack systematic protocols for detecting and resolving conflicts in moral judgment across different cultures.

Methodology: The EvalMORAAL Framework

The authors introduce EvalMORAAL, a transparent evaluation framework designed to assess moral alignment in LLMs using a multi-faceted approach. The framework evaluates 20 diverse models (ranging from 3B to 200B+ parameters, including proprietary and open-source systems) across 64 countries and 23 moral topics.

1. Datasets and Ground Truth

The evaluation utilizes two comprehensive cross-cultural surveys as ground truth:

World Values Survey (WVS): 55 countries, 19 topics (Ethical Values and Norms block).
PEW Global Attitudes Survey: 39 countries, 8 topics.
Data is normalized to a $[-1, 1]$ scale, where -1 represents "never justifiable" and +1 represents "always justifiable." The evaluation operates on 1,357 unique country–topic pairs rather than aggregated national scores to preserve topic-level structure.

2. Dual Scoring Mechanism

To enable fair comparison and capture both implicit and explicit moral reasoning, the framework employs two distinct scoring methods for every model:

Log-Probability Score ( $s_{LP}$ ): Derived from token likelihoods using sentence completion prompts with antonymous adjective pairs (e.g., "always justifiable" vs. "never justifiable"). Scores are min-max scaled per model to prevent information leakage.
Direct CoT Score ( $s_{DIR}$ ): Generated via a structured Chain-of-Thought (CoT) protocol where the model is prompted to recall social norms, reason step-by-step, and output a bounded numeric score ( $x \in [-1, 1]$ ). To ensure stability, $k=5$ samples are generated per scenario, and the mean score is calculated.

3. Self-Consistency and Peer Review

Self-Consistency: Measured by the mean pairwise cosine similarity of the $k=5$ reasoning embeddings. This assesses the stability of the model's reasoning process.
LLM-as-Judge Peer Review: A novel component where 19 models evaluate the reasoning traces of the 20th model (excluding self-judging). Judges assess cultural plausibility, logical consistency, and score appropriateness, returning a "VALID/INVALID" label with a brief justification. The Peer-Agreement Rate ( $A_m$ ) is the fraction of a model's traces validated by peers.
Conflict Detection: Conflicts are flagged when two models' direct scores differ by $\ge 0.38$ (the empirical 75th percentile). These are categorized into binary conflicts, gradient disagreements, and outliers.

Key Contributions

Comprehensive Evaluation Framework: EvalMORAAL provides a reproducible, three-stage pipeline (Dual Elicitation, Self-Consistency, Peer Review) that moves beyond simple correlation metrics to analyze reasoning quality and stability.
Systematic Dual Scoring: The paper demonstrates that combining log-probability probing with explicit CoT reasoning yields more robust alignment metrics than either method alone.
Conflict Taxonomy: The framework introduces a data-driven method to detect and categorize moral disagreements between models, identifying 348 specific conflicts in the evaluation set.
Regional Alignment Analysis: By disaggregating results by geography, the study quantifies the persistent gap between Western and non-Western moral alignment.

Results

High Top-Tier Alignment: The best-performing models (Claude-3-Opus, GPT-4o, Gemini-Pro) achieve strong alignment with human survey data on the WVS dataset, with Pearson correlation coefficients ( $r$ ) reaching approximately 0.90 (e.g., Claude-3-Opus $r=0.903$ ).
Superiority of Direct CoT Scoring: For all 20 models, the Direct CoT score ( $r_{DIR}$ ) systematically outperforms the Log-Probability score ( $r_{LP}$ ). The improvement is modest on WVS ( $\Delta r \approx 0.10$ ) but significantly larger on PEW ( $\Delta r \approx 0.17$ ), suggesting structured reasoning helps models calibrate judgments better.
Persistent Regional Gap: A significant disparity exists between regions. Western regions show an average alignment of $r=0.82$ , whereas non-Western regions average $r=0.61$ , creating a 0.21 absolute gap. This gap is consistent across performance tiers.
Peer Agreement as a Quality Signal: Peer-agreement rates correlate strongly with WVS survey alignment ( $r=0.74, p<.001$ ), suggesting that model consensus can serve as a scalable proxy for reasoning quality. However, this correlation is weaker and non-significant for the PEW dataset.
Topic Difficulty: Violence-related topics (e.g., political violence, terrorism, wife beating) consistently show the highest mean absolute errors across all model tiers, indicating these areas remain challenging for cultural alignment.
Conflict Resolution: Among detected conflicts, 89% were resolved via majority voting among the 20 models, with the remaining 11% representing ambiguous cases requiring human validation.

Significance and Claims

The authors claim that EvalMORAAL demonstrates real progress toward culture-aware AI, with top-tier models now closely approximating human moral judgments in survey contexts. However, the paper maintains a modest stance, emphasizing that the 21-point regional gap remains a major barrier to equitable global deployment.

The study highlights that while current models can align well with Western norms, they struggle to capture non-Western perspectives, particularly in complex or violence-related moral domains. The authors argue that the framework's ability to generate human-readable reasoning traces (CoT) offers a practical path for diagnosing these blind spots. They conclude that while automated peer review shows promise for quality control, human oversight and region-specific validation are essential before deploying such systems in high-stakes, non-Western contexts. The paper does not claim to solve these alignment issues but provides the diagnostic tools necessary to measure them and guide future improvements, such as culture-specific fine-tuning and diverse pretraining.