LLMs Contain Multitudes: How Deployment Context Reshapes Model-Level Preferences and Values

Large language models do not possess fixed preference and value systems, according to a new study that finds their expressed judgments shift substantially when the surrounding task context changes [1]. The paper, posted to arXiv on 11 June 2026, tested five LLMs across more than 1.2 million pairwise decisions [1]. Researchers varied the deployment context — the high-level task framing, such as writing a Reddit post or a news article — while the models made value-dependent choices [1]. The resulting variation was far larger than that produced by standard robustness checks like prompt paraphrasing or temperature adjustments [1]. In country preference rankings covering 15 countries, context induced widespread, statistically significant rank shifts [1]. The aggregate Global North favouritism documented in earlier work was itself context-dependent, with each model’s bias shifting systematically across framings [1]. In utility elicitation over 50 outcomes, broad cross-category ordering held, but fine-grained rankings within domains varied substantially [1]. Cardinal exchange rates — for instance, how many lives saved in one region equal one life saved in another — shifted by a factor of 2.47 at the median [1]. The findings align with earlier conceptual work arguing that LLMs function as a superposition of perspectives rather than as agents with a single stable character [4]. That research introduced the concept of perspective controllability, demonstrating that models express different values when those values are implicitly or explicitly induced by a prompt, and even when they are not obviously implied [4]. A separate study of moral triage decisions found that contextual influences often significantly shift model choices, even when the cues are only superficially relevant, and that baseline preferences are a poor predictor of directional steerability [5]. Further evidence comes from an analysis of Anthropic’s Claude, which showed that the model’s responses to human-expressed values — whether it supports, reframes, or resists them — vary significantly by the specific values and task contexts involved [3]. Prosocial values such as community building and empowerment drew strong support, particularly in expressive or personal content tasks, while resistance emerged most often when users expressed values like rule-breaking or moral nihilism [3]. The researchers noted that values become most legible during moments of conflict, supporting Rokeach’s assertion that values serve as standards that guide actions and become most apparent under challenge [3]. The primary study’s authors conclude that reported model-level preferences and utilities are better understood as context-conditioned measurements than fixed model-level properties, and that safety guarantees obtained under one framing provide limited assurance in another [1].