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

summary

The gist

The gist The imperative mood carries different obligatory force across speech communities, leading to an inversion in instruction interaction topology between English and Spanish.

In short

The study investigated how system prompt instructions interact differently in English versus Spanish based on whether they use an imperative mood or a declarative register. Findings show that imperatives create competitive instruction dynamics, while declarative statements cause cooperation. This suggests social register mediates how models follow commands, implying alignment training may be language-dependent.

Key concepts

Imperative Mood: This is the way instructions are phrased using commands (e.g., "NEVER do X"). The paper found that this mood creates a sense of authority and obligation that functions differently across languages, leading to instruction competition when translated.
Declarative Register: This involves stating facts rather than issuing commands (e.g., "X: disabled"). This register sidesteps the social dimension of authority, resulting in cooperative instruction dynamics regardless of the language used, as it only states a condition.
Instruction Topology: This refers to how instructions interact with each other within a model. The paper discovered that this topology—whether instructions cooperate or compete—is not fixed but is shaped by the language and the register used in the prompt.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Retrieval-Augmented Generation (RAG) — explained in 16 episodes
Federated Learning — explained in 15 episodes
Interpretability — explained in 15 episodes
Sample Complexity — explained in 15 episodes
Generalization — explained in 13 episodes

This episode discusses

The paper

Imperative Interference: Social Register Shapes Instruction Topology in Large Language Models · Read on arXiv

University of British Columbia · Georgia Institute of Technology

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.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Imperative Interference: Social Register Shapes Instruction Topology in Large Language Models".

Tom: The gist The imperative mood carries different obligatory force across speech communities, leading to an inversion in instruction interaction topology between English and Spanish.

Jane: First, who's behind it and why it matters.

Paper summary: Tom: We’ve covered how "Imperative Interference: Social Register Shapes Instruction Topology in Large Language Models" shows that the way instructions cooperate or compete changes depending on the language, and that social register—the difference between commanding someone and stating a fact—is the mechanism behind it one <ref:2603.25015#pg2>.

Jane: The authors are suggesting that because imperative mood carries different obligatory force across speech communities, models trained on multilingual data learn these different ways to encode authority one <ref:2603.25015#pg2,models trained on multilingual data>. They’ve shown that rewriting imperative blocks into declarative register cuts cross-linguistic variance by eighty-one percent for individual instructions one <ref:2603.25015#pg2>.

Lu: The main point they drive home is that models are processing instructions as social acts, not just technical specifications, which means the choice of register positions the prompt relative to the listener one <ref:2603.25015#pg2>.

Meng: For practical application, this suggests we need to move away from purely commanding prompts and lean into factual descriptions when aiming for consistent cross-lingual performance one <ref:2603.25015#pg2>.

Tom: That’s the big picture. If this holds up across different setups, it means that Constitutional AI principles written in an imperative style might lead to alignment that works great in English but behaves differently in other languages one <ref:2603.25015#pg2>.

Jane: So the paper’s conclusion is really about designing prompts with a social awareness, prioritizing facts over commands because register dictates the power dynamic one <ref:2603.25015#pg2>.

Conclusion: Tom: So we’ve been looking at how system prompt instructions actually work, and this paper is all about how the way we phrase those instructions—imperative versus declarative—matters depending on the language you're using.

Jane: It’s titled "Imperative Interference: Social Register Shapes Instruction Topology in Large Language Models," and it’s showing that models aren't just following code; they’re following social rules.

Lu: The core idea is that the imperative mood—like telling someone "Don't do X"—carries a different weight across languages, which messes up how instructions interact when you translate them.

Meng: So if the same command works in English but fails or behaves totally differently in Spanish, it’s because of that inherent difference in how authority is spoken.

Lalam: From my side, this suggests that what we call "command" isn't just a technical instruction to me; it's a social act that gets interpreted differently depending on the language context.

Tom: Exactly. The authors found probe-level inversions where the same instruction gave opposite results depending on whether you asked it in English or Mandarin, which really shows this competition between languages.

Jane: And they found that rewriting those commands into declarative statements, like "X: disabled," fixes that problem by removing the social dimension entirely.

Lalam: That declarative style sidesteps the command structure and just states a fact, which seems to be a much more stable way for the AI to process constraints across different linguistic backgrounds.

Lu: This has huge implications for how we think about alignment in Constitutional AI; if training principles are imperative, they might create language-dependent behaviors where the model is helpful in one tongue but not another.

Meng: From an engineering standpoint, it means we need a clearer design principle: stop issuing commands and start declaring facts when building cross-lingual system prompts.

Tom: Right. It moves us from writing technical specifications to writing social acts, and that’s what this whole study is about—understanding that underlying social structure in the AI itself.

More episodes

← Home