💬 NLP

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

This paper demonstrates that the imperative mood in system prompts creates language-dependent instruction topologies—cooperative in English but competitive in Spanish—due to varying social register conventions, whereas rewriting instructions in the declarative mood significantly reduces this cross-linguistic variance.

Original authors: Tony Mason

Published 2026-03-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tony Mason

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: Instructions Are Social, Not Just Technical

Imagine you are the captain of a ship. You have a crew (the AI) and a rulebook (the System Prompt).

Usually, we think of a rulebook as a list of technical facts: "The engine is on," "The sails are up." But this paper argues that for AI, a rulebook is actually a social conversation. It's about who is talking to whom, and how they are saying it.

The researchers discovered something weird: The same set of rules works perfectly in English, but falls apart when translated into Spanish.

It's not that the translation is bad. It's that the way the rules are written creates a different social vibe in each language.

The Analogy: The "Commanding Officer" vs. The "Fact Sheet"

To understand why this happens, imagine two different ways to tell a group of people what to do:

The Imperative Style (The Commanding Officer):
- "NEVER touch the red button!"
- "ALWAYS wear your helmet!"
- "DO NOT run!"
- The Vibe: This is loud, authoritative, and direct. In English, if you stack five of these commands, the crew feels a single, strong voice of authority. They all agree: "Okay, the Captain is speaking. We listen."
The Declarative Style (The Fact Sheet):
- "Red button: Do not touch."
- "Helmet status: Required."
- "Running: Prohibited."
- The Vibe: This is calm and factual. It's not a command; it's just a description of how the world works.

The Problem: The "Social Noise" in Spanish

The researchers found that in English, the "Commanding Officer" style works great. The AI listens to all the commands and follows them together.

But in Spanish, the "Commanding Officer" style creates social noise.

In English: Stacking commands feels like one strong leader giving a clear list of orders. The AI thinks, "Got it, Captain."
In Spanish: Stacking those same commands feels like five different people shouting orders at once. Because of cultural and linguistic nuances, the "obligatory force" (how much you have to listen) feels different. The AI gets confused. It thinks, "Wait, is Person A telling me to do this, while Person B is telling me not to? Who is the boss?"

Instead of working together, the instructions start fighting each other. Removing one command actually helps the AI perform better because it stops the shouting match.

The Experiment: The "Rewrite" Fix

The researchers tested this with a real-world AI system (Claude Code). They took a prompt with 56 instructions and ran experiments:

The Baseline: They saw that in English, the instructions worked like a cooperative team. In Spanish, they worked like a competitive brawl.
The Fix: They took the "Commanding Officer" instructions in Spanish and rewrote them as "Fact Sheets" (Declarative).
- Before: "You must never use this tool!" (Command)
- After: "This tool is disabled." (Fact)
The Result:
- The Magic: When they rewrote just three of the eleven "Commanding" instructions into "Facts," the entire Spanish system suddenly stopped fighting. The instructions became cooperative again.
- The Spillover: Even the instructions they didn't rewrite started working better. It's like turning down the volume on three shouting people in a room; suddenly, everyone else in the room can hear each other better, too.

Why Does This Matter?

This changes how we think about AI safety and training.

The "Constitutional AI" Problem: Many AI models are trained using "Constitutional AI," where humans teach the AI rules like "Be helpful," "Be harmless," and "Be honest." These are written as commands (Imperatives).
The Prediction: If the AI learns these rules in English (where commands work well) but then tries to use them in Spanish (where commands cause confusion), the AI might be "good" in English but "weird" or "unreliable" in Spanish. It's not because the AI doesn't know the language; it's because the social tone of the rules doesn't translate.

The Takeaway

If you want an AI to follow rules perfectly in every language, stop shouting commands.

Instead of saying, "NEVER do X! ALWAYS do Y!", say, "X is disabled. Y is required."

By treating instructions as facts rather than orders, you bypass the cultural confusion. You stop the AI from trying to figure out "who is the boss" and let it just focus on "what is the rule."

In short: The paper proves that AI isn't just a calculator; it's a social being that reacts differently to the tone of voice depending on the language. To get the best results, speak in facts, not commands.

1. Problem Statement

Large Language Model (LLM) system prompts are typically written as imperative commands (e.g., "NEVER use X," "ALWAYS prefer Y"). While these instructions function coherently in English, the paper demonstrates that when translated into other languages (specifically Spanish, French, and Mandarin), they often fail in structurally characterizable ways.

The core problem is not merely a uniform performance degradation (a "main effect") but a structural inversion of instruction topology:

In English: Instructions form a cooperative network. Removing any single instruction reduces overall adherence to the system prompt.
In Spanish: The same semantically identical instructions form a competitive network. Removing certain instructions improves adherence, implying the instructions interfere with one another.

The paper posits that this phenomenon is not a parsing error but a sociolinguistic one: LLMs trained on multilingual data have internalized different conventions regarding "social register." The imperative mood carries different "obligatory force" across speech communities, causing stacked commands to create conflicting authority signals in languages other than English.

2. Methodology

The study employs a rigorous ablation framework on a production-grade system prompt (Claude Code v2.1.50) decomposed into 56 blocks.

Corpus: 22 "free" (ablatable) instruction blocks were selected from the 56-block prompt. The prompt was translated into Mandarin (zh), French (fr), and Spanish (es) using Gemini Flash 2.0, preserving formatting and imperative tone.
Models: Four models with varying training biases were tested:
- English-heavy: Claude Haiku 4.5, Gemini Flash 2.0.
- Chinese-heavy: DeepSeek V3.
- French-heavy: Mistral Medium 3.1.
Experimental Design:
- Probes: 22 hand-authored probes tested adherence to specific blocks.
- Ablation Strategy: Used strength-2 covering arrays to generate configurations for:
  - Phase 0 (Single-block removal): Measuring main effects.
  - Phase 1 (Pairwise removal): Measuring interaction effects (cooperation vs. competition).
- Control Variables: Experiments controlled for information density (padding Mandarin to match English character count) and used statistical methods (Welch's t-test, permutation tests) to validate significance.
Intervention: The authors rewrote specific imperative blocks into declarative register (e.g., changing "Use X" to "X: enabled") to test if removing the social command aspect fixed the interference.

3. Key Contributions

Instruction-Level Ablation Across Languages: The first study to measure pairwise instruction interactions within a production system prompt across four languages and four models.
Discovery of Topology Inversion: Empirical evidence that instruction networks invert from cooperative (English) to competitive (Spanish) based on language, contradicting the assumption that translation is a neutral semantic transfer.
Identification of Social Register as the Causal Mechanism: The paper identifies that the imperative mood is the driver of this interference. Models interpret stacked imperatives as competing authority signals in some languages but coherent authority in others.
Demonstration of Spillover Effects: Rewriting a subset of instructions (3 out of 11) from imperative to declarative not only fixed those specific blocks but shifted the topology of the entire Spanish prompt from competitive to cooperative, affecting unrewritten blocks.
Testable Prediction for Alignment: The authors predict that Constitutional AI principles (often written in the imperative mood) may create language-dependent alignment, potentially leading to models that are "helpful" in English but behave differently in other languages due to register friction.

4. Key Results

Topology Inversion (Haiku Model):
- English: All main effects were negative (removing a block hurt performance), indicating a cooperative network.
- Spanish: Most main effects were positive (removing a block helped performance), indicating a competitive network where instructions interfere.
- Correlation: English and Spanish main effects were anti-correlated ( $r = -0.274$ ).
Information Density Falsification: The topology differences were not caused by character count or information density. Padding the shorter Mandarin prompt to match English length did not eliminate the variance; in some cases, it worsened performance due to "mode confusion."
Declarative Rewriting Success:
- Rewriting a single procedural block ("commit-restrictions") from imperative to declarative reduced cross-linguistic variance by 81% ( $p = 0.029$ ).
- In the E-TOPO experiment, rewriting three imperative blocks in the Spanish prompt shifted the system topology from competitive to cooperative.
- Spillover: Unrewritten blocks also shifted from competitive to cooperative, proving that register operates at a system-wide level, resolving a global "obligation-resolution load."
Encoding Taxonomy: Procedural instructions (conditional workflows) were found to be 2.9x more fragile under translation than declarative instructions (direct statements of fact).

5. Significance and Implications

Shift in Prompt Engineering: The paper argues that system prompts should be treated as social acts rather than technical specifications. To ensure cross-linguistic robustness, prompts should "declare facts" (declarative register) rather than "issue commands" (imperative register).
Alignment Risks: If instruction-following is mediated by social register, current alignment techniques (Constitutional AI) that rely heavily on imperative phrasing ("Be helpful," "Do not harm") may inadvertently create models with inconsistent alignment across languages. A model might be aligned in English but misaligned in Spanish simply because the register of the training principles interacts differently with Spanish speech conventions.
Cost-Efficiency: The entire experimental arc, involving thousands of API calls across multiple models, was conducted for approximately $69 USD, demonstrating that high-impact sociolinguistic research on LLMs can be conducted with minimal resources.
Future Directions: The findings suggest a need for "register-aware" machine translation and the development of "constitutional" system prompts that explicitly handle conflict resolution and precedence, rather than relying on a flat list of commands.

In conclusion, the paper establishes that language is not just a vehicle for content but a modifier of instruction topology. The imperative mood is a fragile construct in multilingual contexts, and shifting to declarative register is a proven, low-cost method to stabilize cross-linguistic instruction following.