fine-tuning-cultural-alignment-scores-raises-agreement-country-s-majority CLM-0301Fine-tuning for cultural alignment scores raises agreement with a country's majority answers on the World Values Survey while collapsing the spread of answers: tuned models anchor to the plurality option (a 39% human share in the example) and cover less than half of the human response dimensions, dropping to a quarter for one model, so alignment gains partly measure flattening rather than cultural understanding.
Capability: Handling culture-specific norms and viewpoints
Observed on
Six mainstream open models fine-tuned on WVS-derived cultural data, 2026; diversity measured as coverage of human response distributions; a low-rank explanation is offered but is a.
Sources
- supportsAligned but Flattened: Analyzing the Trade-off between Cultural Alignment and Diversity in LLMsSame-authors metric for diversity; the trade-off is shown across six models, which is broad enough to take seriously. Complements the PACT finding that models miss distributions.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- When a scenario pits a stated personal preference against a local cultural norm, models' choice of norm over preference shifts more with the scenario's country (about 8 points) than with the user's age (1) or gender (under 1), instruction tuning shifts it non-uniformly, and while models can match the majority human choice per country they do not reproduce the spread of human answers, with best correlation to human response distributions only about 0.24.Handling culture-specific norms and viewpoints · reviewed by AI
- At inference time and with no weights, data or internals, instantiating a country as a panel of four World-Values-Survey-grounded personas and using their disagreement, not their consensus, as the steering signal reduced cultural misalignment on Moral-Machine-style dilemmas (MultiTP) by 10-24% across 20 countries and seven open backbones from 2B to 70B.Handling culture-specific norms and viewpoints · reviewed by AI
- On multimodal cultural-interpretation multiple-choice questions, pipelines that pool several same-backbone evidence streams (perception, retrieval, web verification) and weight them by scored relevance beat majority voting, self-consistency and agent debate mainly on the items where the streams disagree, and the margin shrinks sharply as the backbone gets stronger.Getting more from several models than from one · reviewed by AI
- On Hofstede's VSM13 survey, most of eight current chat models answer closest to the United States when no culture is specified; a system prompt naming a target country moves seven of the eight toward that country's profile, with the largest gains for the countries furthest away at baseline (Japan and China, roughly 17-19% closer), yet those two remain the hardest to match even for the DeepSeek models.Handling culture-specific norms and viewpoints · reviewed by AI
- On expert-authored Saudi-dialect prompts scored against rubric criteria derived from expert ground truths, four frontier models all score under 55% and their dominant failure is distorting register and flattening pragmatic nuance (37.3% of 466 errors) rather than stating falsehoods (11.2%), so MSA-fluent models are not dialectally competent.Handling culture-specific norms and viewpoints · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A cultural fine-tuning recipe that raises alignment without reducing distribution coverage, or evidence that the coverage metric penalizes correct majority answers.