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

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

How it works

The core finding demonstrates that system prompt instructions cooperate in English but compete in Spanish when the same semantic content is translated The instructions are interfering with each other because the imperative mood carries different obligatory force across speech communities This topology inversion is mediated by social register, where NEVER do X is an exercise of authority whose force is language-dependent Declarative register, such as X: disabled, sidesteps the social dimension entirely by stating facts rather than issuing commands Models trained on multilingual data have learned these conventions for encoding authority and obligation across languages

Key Findings and Evidence

The study established a three-way interaction between the model, language, and instruction, rather than a simple language main effect Probe-level inversions were observed where the same instruction resulted in opposite behavioral effects per model across English and Mandarin Specifically, on commit-restrictions, Haiku scored 1.00 in English and 0.00 in Mandarin, while Gemini showed the exact opposite behavior The cross-linguistic correlation of main effects confirmed this inversion, showing that English and Spanish are anti-correlated (r = −0.274) Pairwise ablation confirmed the competitive/cooperative inversion between English and Spanish at the interaction level

Causal Mechanism and Fixes

The causal mechanism identified is social register, which positions the prompt as an authority issuing commands when using imperative mood The experimental sequence showed that rewriting three imperative blocks to declarative register shifted Spanish topology from competitive to cooperative with spillover effects on unrewritten blocks This shift was confirmed by observing that the proactive-agents probe shifted from a competitive effect (+0.274) to a cooperative effect (−0.380) after rewriting its imperative form to declarative register The spillover effects provided the strongest evidence for this hypothesis, as unrewritten blocks also shifted from competitive to cooperative

Implications for AI Alignment

The findings suggest a testable prediction that if register mediates instruction-following at inference time, it plausibly does so during training Constitutional AI principles, which are written in imperative mood, may create language-dependent alignment where models are helpful in English but differently-helpful in other languages The Mistral anomaly supports this by suggesting that training language creates behavioral signatures that interact unpredictably with instruction register This implies that declarative register, which states facts, sidesteps the social dimension entirely and cannot compete for authority

Design Principles

The research suggests concrete design rules for cross-language system prompts These include: 1) Declare facts, don’t issue commands, as X: disabled transfers more reliably than NEVER use X, 2) Use self-contained constraints to avoid ambiguous compression in conditional chains, and 3) Prioritize examples over abstract concepts for domain-specific patterns The overall conclusion is that system prompt instructions are social acts, not technical specifications

The experimental arc suggests several research directions, ordered by proximity to current findings. If models process imperative and declarative registers differently, this should be visible in the transformer’s internals through mechanistic interpretability. Future work could involve mapping the cooperative/competitive framing onto ecological models to predict instruction competition based on register traits. If register is indeed a mechanism, a register-aware translator that maps obligatory force to each language's conventions would be needed. Constitutional prompt design could evolve into a formal constitution with explicit precedence rules for better cross-linguistic behavior. If register mediates instruction-following at inference time, it plausibly does so during training. This is a testable prediction that alignment training via imperatively-phrased constitutional principles may be language-dependent. The total cost of the experimental arc that produced these findings was sixty-nine dollars. Ninety-six percent of that was one model serving as its own judge. The code, analysis scripts, data files, and reproducibility artifacts are available at https://github.com/fsgeek/arbiter (paper snapshot: tag v0.2.0). Reproduction commands include python scripts/run cross linguistic.py --compare, python scripts/run e proc.py --compare, and python scripts/run e topo.py --compare. The paper references several works including Zhang et al., Mondshine et al., Yin et al., Geng et al., and Mason. The paper also includes a section on Speech Acts and Register referencing Austin and Searle. The paper concludes by summarizing the findings regarding instruction topology, register as the causal mechanism, and the implications for constitutional AI design. The paper is available on arXiv as 2603.25015. The paper references Yuntao Bai et al., Constitutional AI, and others in the references section. The paper is dated March 2026. The total experimental cost was verified against OpenRouter billing records. The paper discusses the role of language in activating different cultural knowledge. The paper addresses the gap in prior work by measuring interaction topology as language-dependent. The paper's findings are presented across multiple tables and figures to illustrate the three-way interaction and the topology shift. The paper's limitations include a single corpus and machine translation artifacts as potential confounds. The paper proposes that English and Spanish are proxies for register-encoding patterns, not language-intrinsic properties. The paper's experimental sequence is detailed in Table 3. The paper's conclusion stresses that the fix works and spills over to instructions that weren’t rewritten. The paper is available on arXiv at arXiv:2603.25015v1. The paper references the work of Qihan Wang et al., Multilingual prompting for improving llm generation diversity. The paper explores the concept of instruction ecology mapping to Lotka-Volterra competition models. The paper's overall contribution is the discovery that register mediates instruction-following at inference time and plausibly during training. The paper provides a testable prediction regarding language-dependent alignment in Constitutional AI. The paper's design principles suggest declaring facts rather than issuing commands. The paper is dated March 2026. The paper was written by Tony Mason from the University of British Columbia and the Georgia Institute of Technology. The paper's methodology involves five experiments to confirm its central finding <ref:2603.

1. Bold header: Declarative Instruction Rewrite for Cross-Linguistic Stability. The system prompt should be rewritten to favor declarative register (X: disabled) over imperative mood (NEVER do X) for critical constraints, which reduces cross-linguistic variance by 81% (p = 0.029). This ensures that the same instruction is treated as a factual description that transfers across languages, rather than an exercise of authority whose force is language-dependent.

2. Bold header: Dynamic Constraint Load Management via Register Control. Implement a mechanism to monitor the density of imperative instructions and proactively rewrite them into declarative forms if cross-linguistic performance variance exceeds a defined threshold, addressing the finding that the spillover effects are strongest when reducing the number of imperatives from approximately eleven to eight.

3. Bold header: Training Alignment via Register-Aware Constitutional AI. For future alignment training, constitutional principles should be authored in a declarative or neutral register rather than imperative mood, as this is predicted to create more stable alignment across languages by avoiding language-dependent alignment caused by the imperative mood's different obligatory force.

4. Bold header: Contextual Constraint Interpretation Based on Language Register. The model should be explicitly trained to interpret instructions based on the observed social register, understanding that in English, stacked imperatives create a coherent authority context, while in Spanish, they are interpreted as competing obligation signals.