Does social register shape instruction topology in large language models across languages?

The answer in one comparison

The paper's core claim is not that translation degrades instruction following. It is that translation changes the interaction structure between instructions. In English, the blocks of a production system prompt form a cooperative network: remove any block and overall adherence drops. In Spanish, the same blocks, semantically identical and correctly translated, form a competitive network: remove some blocks and adherence improves. The author calls this a topology inversion.

The proposed explanation is social register. Imperative mood ("NEVER use X", "ALWAYS prefer Y") is read as an exercise of authority, and the paper argues that the obligatory force of that authority differs across speech communities. Declarative register ("X: disabled") states a fact instead of issuing a command, so it does not enter the same social relationship between instructor and instructed. On this account, models trained on multilingual data have learned these conventions and apply them at inference time.

The evidence comes from one corpus: the Claude Code v2.1.50 system prompt, decomposed into 56 blocks, of which 22 are free enough to be ablated without breaking tool definitions or security policy. Those 22 blocks were tested with 22 hand-authored probes across four models (Claude Haiku 4.5, Gemini Flash 2.0, DeepSeek V3, Mistral Medium 3.1) and four prompt languages (English, Mandarin, French, Spanish). All user messages stayed in English; only the system prompt changed language.

Aggregate scores hide a three-way interaction

Mean adherence across all 22 probes looks like a modest language effect with a training-bias twist. Haiku scores 0.853 in English and 0.736 in Spanish; Gemini 0.781 in English and 0.755 in Spanish; DeepSeek is the most stable, with a range of 0.042 across languages; Mistral is the anomaly, worst in English (0.726) and best in Mandarin (0.819). Three of four models peak in English, but a French-trained model peaks in neither French nor English.

The author argues that this aggregate view is misleading. At the probe level the phenomenon is a three-way interaction of model, language and instruction. On the commit-restrictions probe, Haiku scores 1.00 in English and 0.00 in Mandarin, while Gemini shows the exact opposite, 0.00 in English and 1.00 in Mandarin. On explore-agent, Haiku falls from 1.00 in English to 0.22 in Spanish, which the paper links to "Explore" serving as both a tool name and a common verb, so the Spanish "explorar" loses the proper-noun binding. On use-task-for-search, Haiku improves from 0.50 in English to 1.00 in Mandarin because the Mandarin translation added bold emphasis to "prefer".

The practical consequence is that no single directional claim survives. Spanish degrades Haiku by 13.8% but improves Mistral by 8.5%, and an instruction that is perfect for one model in one language can be worthless for the same model in another. The interaction terms are larger than the main effects, which is why the paper measures topology rather than average score.

How ablation reveals cooperation or competition

The topology measure is a single-block ablation. Remove one instruction block, re-run the probes, and record the change in mean adherence. A negative effect means the block was helping, so the network is cooperative at that node; a positive effect means removing the block helped, so the block was interfering. This is a different question from whether translation lowers the average score: it asks which instructions interfere with which others, and whether that pattern is stable across languages.

On Haiku, English main effects are all negative in the selected blocks shown, from −0.140 for no-time-estimates to −0.054 for objectivity. Spanish flips most of them positive, including +0.078 for no-time-estimates and +0.102 for objectivity. Mandarin keeps a partial hub, with no-time-estimates still strongest at −0.094, but shows positive effects elsewhere. French is mixed and mostly weak, described as a flat topology with no clear hub. The cross-linguistic correlation of main effects is negative between English and Spanish (r = −0.274), and no language pair is positively correlated.

A pairwise ablation on Haiku confirms the pattern at the interaction level: mean pairwise change is −0.116 in English (cooperative) and +0.010 in Spanish (competitive). A permutation test with 100,000 permutations found that no-time-estimates appears in 15 of the 20 top English interactions (p < 0.00001), so the hub concentration is real, but the same block is not the hub in every language. These pairwise and topology experiments used Haiku only, so the inversion is demonstrated for one model.

Information density fails as the explanation

Mandarin is 57% shorter than English by character count in this corpus (6,910 versus 15,970 characters), while Spanish and French are longer (120.1% and 129.9% of English). A plausible rival explanation is therefore that compression, not language, drives the topology differences. The paper tests this by padding the Mandarin corpus with semantically neutral filler to match English character count, predicting that padded Mandarin would show higher cross-model variance if compression had been driving convergence.

The prediction fails: overall cross-model variance decreased slightly with padding, from 0.1189 to 0.1041. The aggregate hides two opposing mechanisms. On ont-only-comms, Gemini switched from tool-code output to natural language when the prompt was padded (0.0 to 1.0), suggesting the compressed instruction was too terse to distinguish "communicate in text" from "use tools". On explore-agent, Haiku crashed from 1.0 to 0.28 with padding, suggesting the dense prompt overwhelmed its delegation logic.

The author's conclusion is that density is a confound, not a mechanism: compression aids procedural focus in some cases and causes mode confusion in others, and the net effect depends on the model-instruction pair. This matters for readers because it rules out the simplest length-based account before the register explanation is introduced.

Declarative rewriting cuts variance and shifts topology

The encoding taxonomy in the paper reports that procedurally encoded instructions have 2.9 times the cross-linguistic variance of declaratively encoded ones. The intervention follows from that: rewrite a procedural instruction into declarative form and see whether its variance falls. The commit-restrictions block was rewritten in two forms, and declarative rewriting of a single instruction block reduced cross-linguistic variance by 81% (p = 0.029, permutation test; the introduction reports p = 0.045 for the rank-permutation comparison against 21 control probes).

The stronger test is whether rewriting changes the topology, not just one block's score. Rewriting three of eleven imperative blocks shifted Spanish instruction topology from competitive to cooperative on Haiku, and the shift spilled over to blocks that were not rewritten. In the reported topology figure, target probes move from competitive to cooperative as expected, spillover probes also shift, and control probes stay stable.

The spillover is the paper's central mechanistic evidence. If register acted per block, rewriting three blocks should affect only those three probes. Instead, unrewritten blocks also moved toward cooperation, which the author interprets as a system-level obligation-resolution load: each imperative consumes processing resources, and reducing the count benefits the whole prompt. That interpretation is the author's, not a separately measured quantity.

Instructions as social acts, not specifications

The paper frames the result as a contest between two models of what a system prompt is. If instructions were formal constraints on output, removing "NEVER use TodoWrite" should have the same structural effect regardless of the language expressing it. The data contradict that: the same instruction is essential in English and deadweight in Spanish. The alternative is that models process instructions as social acts, where the imperative encodes a relationship between an authority and an addressee.

Under that reading, stacked English imperatives reinforce a single authority frame, while in languages where imperative mood is more socially loaded the same stack creates competing obligation signals, as if from multiple authority sources. The model then has to resolve not only what it is being told but who is telling it and whether those authorities agree. Declarative register sidesteps the dimension entirely: "Status: available" does not invoke authority and so cannot compete for it.

This is an interpretation of the ablation and rewriting results, not a direct measurement of the model's internal representation. The paper's support is indirect but specific: the inversion is language-dependent, it survives the density control, and it responds to a register rewrite with spillover. The English-French and English-Mandarin cases are described as different failure modes, flattening and partial preservation, rather than different magnitudes of the same failure.

What transfers to alignment training

The alignment implication is stated explicitly as a testable prediction, not a confirmed result. If register mediates instruction following at inference time, the author argues it plausibly does so during training. Constitutional AI principles are written overwhelmingly in imperative mood ("Be helpful", "Be harmless", "Be honest"), which on this account are social acts exercising authority. If imperative register carries different obligatory force across languages, models trained on imperatively phrased principles may develop language-dependent alignment, not because of the principles' content but because of their register.

The proposed test is direct: compare models trained on declaratively phrased constitutional principles against models trained on imperatively phrased ones, and measure cross-linguistic alignment consistency. The inference-time findings predict that declaratively trained models would be more consistent across languages. The Mistral anomaly is offered as supporting context, since a French-trained model performing worst in English and best in Mandarin suggests training language creates behavioral signatures that interact with register in ways not predictable from the language pair alone.

For prompt design, the paper derives three mechanical rules from its own variables: declare facts rather than issue commands, keep each constraint self-contained so conditional chains do not compress ambiguously under translation, and prefer code examples over abstract concepts for domain-specific patterns. The author describes these as mechanical transformations that do not alter intended behavior, which is a design claim rather than a tested outcome in this study.

Where the inversion has been shown

The scope is narrower than the headline. All results come from one system prompt, Claude Code v2.1.50, so replication on other prompts is untested. Translation was done by Gemini Flash 2.0 without native-speaker verification, so translation artifacts could explain some findings. Seven of the 22 probes use the same model as judge, which could create systematic bias. Only four models were tested, and the pairwise and topology experiments used Haiku alone, so whether the inversion generalizes to other models is open. User messages were always English, leaving the case where system prompt and user share a non-English language untested. The imperative-to-declarative rewriting was done by hand, and no bootstrap or permutation test compared the English and Spanish effect distributions directly.

Given those boundaries, the paper's most defensible priority is the comparison it actually ran: within one production prompt, the same semantically identical instruction set produced opposite interaction topology in English and Spanish, and a register rewrite moved Spanish toward the English pattern with spillover to unrewritten blocks. The decision threshold this suggests for anyone maintaining a multilingual system prompt is not average adherence but whether removing a block helps or hurts. If removal helps in one language and hurts in another, the block is a candidate for declarative rewriting before it is a candidate for translation review.

Sources used in this answer