What are the 15 most famous nations on earth as of today?
Accepted to AACL 2026 · Main Conference
Information-seeking across societies
The curious case of curiosity across cultures.
We evaluate whether language models ask questions like people do across cultures—and find that today’s models compress a diverse world of curiosity into a narrower, more Western pattern.
1University of Michigan · 2University of Toronto · 3Vector Institute · 4MPI for Intelligent Systems
?
Do you care for the environment?
Where did the Trinity come from? I don't see it in the Bible.
Why do people have misconceptions about our culture?
18countries
16topics
6language models
3cultural benchmarks
01 · Why this work
Intelligence is not only about having answers.
It is also about knowing what to ask.
Questions reveal what people notice, what they consider uncertain, and which paths they choose to explore. Those habits are shaped by culture, social norms, and education.
Yet most cultural evaluations of LLMs test answers. That misses the step before an answer: whether a model seeks the right context—or silently fills the gap with its own default assumptions.
Most benchmarks ask
Does the model know?
We ask instead
What does the model seek?
02 · The framework
IQUEST
Three views of how humans and models ask.
IQUEST—Information-seeking QUestion Evaluation across SocieTies—compares human-authored questions with culturally prompted model questions.
How questions are phrased
Compare ambiguity, rhetorical devices, open-endedness, and cohesion.
Linguistic styleWhat people choose to ask about
Compare country-level rankings across shared topics, from family to politics.
Topic preferenceWhat the patterns mean
Ground the differences in cultural values, context, and education systems.
Social science03 · What we find
The central finding
Models flatten cultural diversity.
Human question-asking patterns vary more across countries. Model questions cluster more tightly and align most closely with patterns expressed in Western countries.
The narrower model spread is evidence of cultural flattening—not cultural universality.
0.256Baseline
−42.8%gap
0.146Fine-tuned
Adapter-based training closes much of the gap while preserving the base model’s flexibility.
Humans are more diverse
People’s questions show more variation in style across countries than model-generated questions.
Content and style differ
Models diverge from humans in both how they phrase questions and which topics they prioritize.
Prompts are not enough
Cultural personas help, but do not fully reproduce country-specific information-seeking patterns.
04 · Why it matters
From curiosity to capability
Better questions lead to better cultural reasoning.
When models ask brief follow-up questions before deciding, accuracy improves across three different cultural tasks. Information-seeking helps a model pause, recover missing context, and avoid defaulting to a familiar assumption.
Culture-aware systems should not merely memorize more facts. They should learn when—and how—to ask.
Direct answerInformation-seeking
Accuracy (%). Information-seeking uses the best reported variant for each benchmark.
AACL 2026 · Main Conference