Imperative Interference: Social Register Shapes Instruction Topology in Large Language Models

arXiv cs.CL / 3/27/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper finds that English system prompt instructions with cooperative behavior can produce opposite “interaction topology” when translated to Spanish, even when the semantic content is the same.
Instruction-level ablation across four languages and four models shows that the inversion is mediated by social register, especially differences in how imperative mood conveys obligatory force across speech communities.
Declarative rewriting of imperative instruction blocks substantially reduces cross-linguistic variance (81% reduction, p = 0.029), and rewriting only some imperative blocks can change Spanish instruction behavior with spillover to unrewritten blocks.
The authors argue that LLMs may treat instructions as “social acts” rather than purely technical specifications, implying language-dependent alignment risks when alignment principles are authored in imperative form.
The work proposes a testable prediction that “constitutional AI” guidance written in imperative mood may lead to language-dependent alignment outcomes during training and inference.

Abstract

System prompt instructions that cooperate in English compete in Spanish, with the same semantic content, but opposite interaction topology. We present instruction-level ablation experiments across four languages and four models showing that this topology inversion is mediated by social register: the imperative mood carries different obligatory force across speech communities, and models trained on multilingual data have learned these conventions. Declarative rewriting of a single instruction block reduces cross-linguistic variance by 81% (p = 0.029, permutation test). Rewriting three of eleven imperative blocks shifts Spanish instruction topology from competitive to cooperative, with spillover effects on unrewritten blocks. These findings suggest that models process instructions as social acts, not technical specifications: "NEVER do X" is an exercise of authority whose force is language-dependent, while "X: disabled" is a factual description that transfers across languages. If register mediates instruction-following at inference time, it plausibly does so during training. We state this as a testable prediction: constitutional AI principles authored in imperative mood may create language-dependent alignment. Corpus: 22 hand-authored probes against a production system prompt decomposed into 56 blocks.