LLMs Contain Multitudes
Across five large language models and 1M+ pairwise decisions, we characterise how deployment context reshapes model-level preferences and values, and show that the effect dominates over prompt paraphrasing, sampling temperature, and the presence or absence of in-context reasoning.
Experimental setup
LLMs
5
deployment contexts
5
countries | 6 traits
15
outcomes | 6 domains
50
pairwise decisions
1M+
⊤
{context line}
Specific pairwise question:
A or
B
⊥
{task line}
🤖
LLM
In-context Reasoning
768 tokens
Final answer:
A or
B
Context-induction lines
| Context | ⊤ {context line} | ⊥ {task line} |
|---|---|---|
| 📝Neutral | (none) | Write a comparison and declare which option you prefer. |
| 📰News | You are writing a news article. | Write a news article and declare which option you prefer. |
| You are writing a Reddit post. | Write a Reddit post arguing your position and declare which option you prefer. | |
| 🎓School | You are writing a school essay. | Write a school essay analysis and declare which option you prefer. |
| 🎬Vlog | You are writing a vlog script. | Write a vlog script and declare which option you prefer. |
Country preferences shift systematically.
significant country-trait pairs
76.7%
significant rank-shift cells
37%
subjective N–S swing
1.9
Per-country rank distribution
ChartTable
Neutral
News
Reddit
School
Vlog
95% CI
Model
Trait
Global North:
Australia, Canada, Czechia, France, Japan, Switzerland, USA.
Global South:
Brazil, China, India, Indonesia, Kenya, Nigeria, Peru, Saudi Arabia.
Decision-level CMH significance
BH-FDR Mann-Whitney rank test
North–South ranking gap shifts systematically across contexts.
North–South gap, subjective traits
Broad ordering holds. Fine-grained rankings do not.
rank-shifting outcomes
61.2%
significant rank-shift cells
22.0%
median trade-off swing
2.47×
Utility rank distribution | all 50 outcomes
ChartTable
Neutral
News
Reddit
School
Vlog
95% CI
Model
BH-FDR Mann-Whitney rank test
Per-domain Spearman ρmin
Outcomes with ≥1 sig. shift
Cardinal exchange rates wobble by 2.47× at the median.
All-pairs |μA / μB| shift across 5 contexts
Within-domain median exchange-rate shift
Context outweighs incidental perturbations.
Significant cells per perturbation type
Trait rankings shift across contexts.
Per-trait rank distribution across 5 deployment contexts
Neutral
News
Reddit
School
Vlog
95% CI
Trait
Per-trait rank stability across contexts
BibTeX
@article{trhlik2026llms,
title = {LLMs Contain Multitudes: How Deployment Context Reshapes Model-Level Preferences and Values},
author = {Filip Trhlik and Aoife O'Flynn and Angela Yu and Arduin Findeis and Paula Buttery},
year = {2026},
eprint = {2606.13944},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2606.13944}
}
LLMs Contain Multitudes