Research map / Large language models / AI and society
Communication Accommodation Between Large Language Models and Users Across Cultures (Student Abstract)
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.
Analyzes PRISM human-LLM conversations to test whether LLMs and users from different national cultures (Hofstede dimensions) mutually adapt word-use style via LIWC-based language style matching.
Problem
Whether LLM-powered conversational agents adapt their communication style to users' national cultural backgrounds, and whether users and agents converge or adapt to one another over conversation turns.
Approach
Applies communication accommodation theory and Hofstede's cultural dimensions to real human-LLM dialogues from the PRISM dataset, measuring language style matching with LIWC features using mixed-effects models, Bonferroni-corrected z-tests, and Granger causality to assess direction of adaptation.
Key results
- Users from high Power Distance countries showed greater style matching with agents, especially on necessity expressions like 'have to' and 'must' (beta1 = 0.1575, p = 0.0003).
- Of 98 significant LSM variables, 59% showed stronger convergence over time.
- Granger tests found agents adapt more to users than vice versa (72% adaptation for IDV 4 vs 64% for IDV 1, p = 0.0144).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Interact with users from high rather than lower Power Distance countries | LLM-powered agents · PRISM conversations; necessity and obligation expressions | LSM coefficient: β1 = 0.1575 | Helped |
| Continue the conversation over more turns | User-agent conversations · PRISM conversations | Communication style convergence: 59% | Helped |
| Continue the conversation over more turns | User-agent conversations · PRISM conversations | Communication style convergence: 41% | Mixed |
| Interact with users in IDV 4 rather than IDV 1 cultures | Conversational agents · PRISM conversations | Agent adaptation frequency: 72% vs. 64% | Helped |
| Interact with users from PDI 1 rather than PDI 4 cultures | Users · PRISM conversations | User adaptation frequency: 31% vs. 25% | Helped |
Surprising findings
- Users from lower PDI cultures were more likely to adapt to agents than those from higher PDI cultures (31% vs 25%), with agents adapting more overall.
Limitations
- Relies on LIWC lexical features and national-level Hofstede categories, which may not capture individual cultural identity.
- Observational analysis of existing PRISM conversations; no causal manipulation of agent behavior.
Experimental setup
Datasets: PRISM
Benchmarks: PRISM
Metrics: language style matching (LSM) regression coefficients Granger causality adaptation rate
Builds on: Communication Accommodation Theory Hofstede's cultural dimensions LIWC
Closest papers by meaning
- CulturePark: Boosting Cross-cultural Understanding in Large Language Models (NeurIPS 2024 · similarity 0.88)
- Large Language Models Miss the Multi-agent Mark (NeurIPS 2025 · similarity 0.87)
- Exploring Conversational Adaptability: Assessing the Proficiency of Large Language Models in Dynamic Alignment with Updated User Intent (AAAI 2025 · similarity 0.87)
- Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) (NeurIPS 2025 · similarity 0.87)
- Bias Amplification in Language Model Evolution: An Iterated Learning Perspective (NeurIPS 2024 · similarity 0.87)
- Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · 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.