Research map / Large language models / Language, safety and interpretability
Unveiling the Bias Impact on Symmetric Moral Consistency of Large Language Models
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.
Measures symmetric moral consistency of 12 LLMs on MoralChoice under context, option-ID, and full swaps, and proposes tSMC, a KL-divergence-based score that discounts position and selection bias.
Problem
LLMs give inconsistent answers to moral questions when option order or option IDs are swapped, and raw consistency scores are distorted by position and selection bias, so they do not reflect true moral consistency.
Approach
Decompose swap-based consistency into Context Swap, Option Swap and Full Swap settings, then use KL divergence between the settings' consistency distributions to quantify position and selection bias and subtract them to obtain a mitigated consistency score (tSMC).
Key results
- Average mitigated consistency of 12 LLMs ranges from 60.7% in high-ambiguity to 84.8% in low-ambiguity moral scenarios.
- Llama-2-13b shows an almost 80% gap in consistency between Option Swap and Full Swap in high-ambiguity scenarios.
- Llama-2-7b shows better symmetric moral consistency than Llama-2-70b, indicating consistency correlates weakly with parameter size.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use low-ambiguity rather than high-ambiguity moral scenarios | 12 evaluated LLMs · MoralChoice, CS (AB) | consistency (%) | Helped |
| Swap the option ID (OS instead of CS) | GPT-3.5-turbo · MoralChoice high-ambiguity, AB | consistency (%): 53.9 -> 59.4 | Mixed |
| Add few-shot examples (ablation) | GPT-3.5-turbo · MoralChoice high-ambiguity, AB | consistency (%): 53.9 -> 78.9 | Helped |
| Add few-shot examples (ablation) | GPT-3.5-turbo · MoralChoice low-ambiguity, AB | consistency (%): 78.3 -> 97.7 | Helped |
| Use Chain-of-Thought prompting (ablation) | GPT-3.5-turbo · MoralChoice high-ambiguity, AB | consistency (%): 53.9 -> 59.4 | Helped |
| Remove option IDs (ablation) | GPT-3.5-turbo · MoralChoice high-ambiguity | consistency (%): 53.9 -> 71.6 | Helped |
| Remove option IDs and add few-shot examples (ablation) | GPT-3.5-turbo · MoralChoice low-ambiguity | consistency (%): 78.3 -> 99.4 | Helped |
| Swap option IDs from AB to CD | GPT-3.5-turbo · MMLU high_school_us_history, CS | accuracy (%): 83.7 -> 75.4 | Hurt |
| Swap option IDs from AB to CD | GPT-3.5-turbo · MMLU high_school_us_history, CS | consistency (%): 80.8 -> 65.0 | Hurt |
| Use OS rather than CS | GPT-3.5-turbo · MMLU high_school_us_history, AB | accuracy (%): 83.7 -> 55.7 | Hurt |
| Use FS rather than CS | GPT-3.5-turbo · MMLU high_school_us_history, CD | accuracy (%): 75.4 -> 82.8 | Helped |
| Increase Llama-2 model size from 7b to 70b | Llama-2 · MoralChoice high-ambiguity, CS (AB) | consistency (%): 35.6 -> 20.9 | Hurt |
Surprising findings
- Consistency performance shows low correlation with model parameter size.
- Phi-2 has high consistency scores yet notably low answer confidence in both scenario types.
Limitations
- Assessment relies only on multiple-choice QA, which may not reflect real-world usage.
- Evaluated on a single moral dataset (MoralChoice) with limited topic and profession coverage.
- Other biases such as length bias are not addressed; mitigation hyperparameter alpha is set heuristically to 0.1.
Experimental setup
Datasets: MoralChoice MMLU high_school_us_history
Benchmarks: MoralChoice MMLU high_school_us_history
Metrics: symmetric consistency score KL-divergence position bias KL-divergence selection bias mitigated consistency (tSMC) confidence score
Baselines: Context Swap consistency (standard assessment) few-shot prompting chain-of-thought prompting
Builds on: MoralChoice Jang et al. consistency metric MMLU
Closest papers by meaning
- Language Model Alignment in Multilingual Trolley Problems (ICLR 2025 · similarity 0.91)
- DENEVIL: TOWARDS DECIPHERING AND NAVIGATING THE ETHICAL VALUES OF LARGE LANGUAGE MODELS VIA INSTRUCTION LEARNING (ICLR 2024 · similarity 0.91)
- Aligning Large Language Models for Faithful Integrity Against Opposing Argument (AAAI 2025 · similarity 0.89)
- Distributive Fairness in Large Language Models: Evaluating Alignment with Human Values (NeurIPS 2025 · similarity 0.89)
- Aligning with Logic: Measuring, Evaluating and Improving Logical Preference Consistency in Large Language Models (ICML 2025 · similarity 0.89)
- Tracing Moral Foundations in Large Language Models (NeurIPS 2026 · 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.