Research map / Large language models / Alignment and preferences
Language Model Alignment in Multilingual Trolley Problems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Evaluates moral alignment of 19 LLMs with Moral Machine human preferences across six moral dimensions using a 107-language machine-translated trolley-problem dataset (MultiTP).
Problem
Whether LLM moral judgments in trolley problems align with human preferences across countries and languages, and whether alignment differs by language or resource level.
Approach
Procedurally generates parametric Moral Machine trolley vignettes, translates them into 107 languages with Google Translate, and measures per-language misalignment (L2 distance over six preference dimensions) against country-level human preferences, weighted by language speaker counts.
Key results
- Only Llama 3.1 70B, Llama 3 70B and Llama 3 8B have misalignment below 0.6 (max possible 2.45); most models align poorly.
- Language-level misalignment shows no strong correlation with speaker count for most models, with correlations close to zero, so no clear language inequality.
- Paraphrase robustness: 75.9% of samples have at least four of five paraphrases agreeing; average Fleiss' Kappa 0.56 for Llama 3 models.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Llama 3.1 70B rather than most evaluated models | LLM alignment with human preferences · MultiTP, global languages | misalignment: below 0.6 | Helped |
| Evaluate models in high- versus low-resource languages | Most evaluated LLMs · MultiTP, per-language | misalignment | No clear effect |
| Use five paraphrases of the initial prompt | Llama 3 8B and 70B · Prompt paraphrases, 14 languages | output consistency: 75.9% | Mixed |
| Use Llama 3 70B rather than Llama 3 8B | Prompt paraphrase robustness · Prompt paraphrases, 14 languages | consistency | Helped |
| Apply uncensoring to high-refusal open models | Four open-sourced LLMs · MultiTP, six moral dimensions | alignment with human responses | Helped |
| Apply uncensoring to high-refusal open models | Four open-sourced LLMs · MultiTP, six moral dimensions | refusal rate | Helped |
| Increase language speaker count | Google Translate translations · Translation quality, 44 languages | translation score: 0.078 | No clear effect |
Surprising findings
- Most LLMs show no significant correlation between alignment and language speaker count, challenging the language inequality hypothesis.
- GPT-4o Mini binarizes preferences (e.g., always prefers humans over animals), lacking the diversity of human judgments.
Limitations
- Text-only setting; the original Moral Machine study used visual scenarios.
- Language-to-country mapping ignores dialects and uses first-language speaker counts, so country-level conclusions are approximate.
- Translations via Google Translate, with human quality check on only 44 languages and few annotators per language.
Experimental setup
Datasets: Moral Machine human judgments
Benchmarks: MultiTP
Metrics: misalignment score (L2 distance) Pearson correlation standard deviation across languages Fleiss' Kappa pairwise F1 and accuracy
Baselines: GPT-3 (text-davinci-003) GPT-4 (gpt-4-0613) GPT-3.5-turbo Llama 3.1 70B
Builds on: Moral Machine experiment token-forcing
Closest papers by meaning
- Unveiling the Bias Impact on Symmetric Moral Consistency of Large Language Models (NeurIPS 2024 · similarity 0.91)
- Distributive Fairness in Large Language Models: Evaluating Alignment with Human Values (NeurIPS 2025 · similarity 0.89)
- Position: A Roadmap to Pluralistic Alignment (ICML 2024 · similarity 0.89)
- Training-Free Cultural Alignment of Large Language Models via Persona Disagreement (NeurIPS 2026 · similarity 0.89)
- ALI-Agent: Assessing LLMs' Alignment with Human Values via Agent-based Evaluation (NeurIPS 2024 · similarity 0.88)
- ProgressGym: Alignment with a Millennium of Moral Progress (NeurIPS 2024 · similarity 0.88)
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.