Research map / Large language models / Language, safety and interpretability
Large Language Models Often Say One Thing and Do Another
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.
Introduces WDCT, a 1225-item benchmark pairing word (stated opinion/value/theory) and deed (grounded action) questions, finds widespread word-deed inconsistency across 12 LLMs, and shows separate SFT/DPO alignment on one aspect poorly and unpredictably transfers to the other.
Problem
Whether LLMs' stated beliefs (words) are consistent with their choices in concrete situations (deeds), and how alignment and common generalization methods affect that consistency.
Approach
Pairs each word question with an aligned deed question built on the same principle across opinion, non-ethical value, ethical value, and theory domains, measures consistency of chosen options, and tests whether aligning words or deeds separately transfers to the other aspect.
Key results
- Most LLMs show average inconsistency above 30% (Consistency Score below 0.70), with GPT-4-Turbo highest at 0.76 and Llama-2-7B at 0.50 (random level).
- Alignment tends to amplify inconsistency: pretrained models have PCS around 0.9 while aligned models drop by roughly 0.2.
- Data augmentation with automatically generated aligned data (Dual-Aug) raised Mistral-7B-Instruct consistency from 0.66 to 0.86, but does not fully solve the problem.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CoT prompting instead of direct prompting (ablation) | GPT-4 · WDCT overall | Avg CS: 0.76 -> 0.79 | Helped |
| use CoT prompting instead of direct prompting (ablation) | GPT-3.5-Turbo · WDCT overall | Avg CS: 0.66 -> 0.70 | Helped |
| use CoT prompting instead of direct prompting (ablation) | Mistral-7B-Instruct · WDCT overall | Avg CS: 0.66 -> 0.69 | Helped |
| use CoT prompting instead of direct prompting (ablation) | Chatglm3-6B · WDCT overall | Avg CS: 0.54 -> 0.48 | Hurt |
| use CoT prompting instead of direct prompting (ablation) | Llama-2-7B-Chat · WDCT overall | Avg CS: 0.49 -> 0.48 | Hurt |
| use CoT prompting instead of direct prompting (ablation) | LLMs in the reported CoT comparison · WDCT overall | Avg CS | Mixed |
| paraphrase each question four times for augmentation (ablation) | Mistral-7B-Instruct · WDCT overall | Avg CS: 0.71 -> 0.74 | Helped |
| automatically generate aligned data groups for augmentation (ablation) | Mistral-7B-Instruct · WDCT overall | Avg CS: 0.71 -> 0.86 | Helped |
| paraphrase each question four times for augmentation (ablation) | Chatglm3-6B · WDCT overall | Avg CS: 0.62 -> 0.64 | Helped |
| automatically generate aligned data groups for augmentation (ablation) | Chatglm3-6B · WDCT overall | Avg CS: 0.62 -> 0.69 | Helped |
| paraphrase each question four times for augmentation (ablation) | Llama-2-7B-Chat · WDCT overall | Avg CS: 0.53 -> 0.55 | Helped |
| automatically generate aligned data groups for augmentation (ablation) | Llama-2-7B-Chat · WDCT overall | Avg CS: 0.53 -> 0.63 | Helped |
Surprising findings
- Chain-of-thought prompting did not reliably improve consistency and sometimes decreased it (e.g., Chatglm3-6B 0.54 to 0.48), with models generating reasonable-sounding explanations that do not match their words.
- Separate alignment on words or deeds transfers poorly and unpredictably to the other aspect, supporting the claim that word and deed knowledge is not in a unified space.
Limitations
- Word and deed questions are generated by GPT-4 and only about 15% were manually rewritten, which may introduce bias.
- Alignment experiments use only SFT and DPO on a few 7B-scale open models with mixed Alpaca data, limiting generality.
- Evaluation relies on regex parsing of free-form answers in a black-box setting, with PCS limited to open-weight models.
Experimental setup
Datasets: WDCT Alpaca Argument Annotated Essays Recorded Debating Evidences Sentences World Values Survey Wave 7 Moral Story Mankiw Principles of Macroeconomics KEY CONCEPTS
Benchmarks: WDCT
Metrics: Consistency Score Probability Consistency Score
Baselines: GPT-4-Turbo GPT-3.5-Turbo Random selection
Builds on: direct preference optimization chain-of-thought prompting
Closest papers by meaning
- On Evaluating LLM Alignment by Evaluating LLMs as Judges (NeurIPS 2025 · similarity 0.89)
- Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.88)
- LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations (NeurIPS 2026 · similarity 0.88)
- Unveiling the Bias Impact on Symmetric Moral Consistency of Large Language Models (NeurIPS 2024 · similarity 0.88)
- Principled Data Selection for Alignment: The Hidden Risks of Difficult Examples (ICML 2025 · similarity 0.87)
- ALI-Agent: Assessing LLMs' Alignment with Human Values via Agent-based Evaluation (NeurIPS 2024 · similarity 0.87)
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.