Aligned but Flattened: Analyzing the Trade-off between Cultural Alignment and Diversity in LLMs
cs.SI, cs.CL
Submitted: 2026-09-01
Updated: 2026-09-01
Comments: Accepted at EMNLP 2026 (Findings)
Project page: https://aligned-but-flattened.github.io
License: http://creativecommons.org/licenses/by/4.0/
The gist: Cultural fine-tuning has become the de facto paradigm for building culture-aware large language models (LLMs), yet existing optimization exclusively for alignment scores provides an incomplete
Terminology
Abstract
Cultural fine-tuning has become the de facto paradigm for building culture-aware large language models (LLMs), yet existing optimization exclusively for alignment scores provides an incomplete portrait of cultural fidelity by systematically obscuring inherent cultural diversity. This unidimensional evaluation lens prompts a fundamental question: do models genuinely perceive distinct cultural nuances, or do they merely memorize dominant cultural values? To address this, we propose a synergistic evaluation framework that jointly formalizes cultural alignment and diversity. Through extensive benchmarking of six mainstream LLMs on the World Values Survey, this framework uncovers a systematic and critical trade-off: the pursuit of cultural alignment consistently incurs an acute expense of diversity, leading to severe "cultural flattening." Investigating this behavioral shift, we demonstrate that these superficial alignment gains stem from models artificially anchoring to dominant majorities, converging onto a monolithic response pattern that wipes out the heterogeneous distributions inherent to human groups. Crucially, our mechanistic analysis suggests that this diversity collapse is not merely a behavioral anomaly but more likely a structural consequence of the low-rank bias inherent in neural network optimization. Therefore, our findings expose the limitations of current post-training paradigms and call for a shift toward alignment objectives that preserve cross-cultural pluralism.
Sources
- Investigating Cultural Alignment of Large Language Models
- Qwen Technical Report
- A Closer Look at Multimodal Representation Collapse
- Self-Alignment: Improving Alignment of Cultural Values in LLMs via In-Context Learning
- DeepSeek-V3 Technical Report
- CulFiT: A Fine-grained Cultural-aware LLM Training Paradigm via Multilingual Critique Data Synthesis
- SGD and Weight Decay Secretly Minimize the Rank of Your Neural Network
- Transformer Feed-Forward Layers Are Key-Value Memories
- SimBench: Benchmarking the Ability of Large Language Models to Simulate Human Behaviors
- The Low-Rank Simplicity Bias in Deep Networks
- Mistral 7B
- Multiple LLM Agents Debate for Equitable Cultural Alignment
- GPT-4 Technical Report
- Gemma 2: Improving Open Language Models at a Practical Size
- Whose Opinions Do Language Models Reflect?
- Exploring the Limits of Large Language Models as a Reflection of Human Cognition: An Illustration in the Context of Moral Judgement
- Cultural Palette: Pluralising Culture Alignment via Multi-agent Palette
- Neural Collapse versus Low-rank Bias: Is Deep Neural Collapse Really Optimal?
- Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language Models
- Gemini: A Family of Highly Capable Multimodal Models
Related papers
- Linking Scalar-Intensity Language to Structural Polarization with Validated Signed-Network Measures
- Detection and Characterization of Coordinated Online Behavior: A Survey
- Omega-N: Interpretable Structural Node Descriptors and Their Applicability Domain
- Discovery of connectivity-trainability trade-off of IQP Circuits for Hamiltonian Optimization
- Epistemic Substitution: How Grokipedia's AI-Generated Encyclopedia Restructures Authority
- Semantic Communities and Boundary-Spanning Lyrics in K-pop: A Graph-Based Unsupervised Analysis