Spending English-Pivot Compute Where It Helps: Translationese-Aware Allocation Across Languages

Multilingual language models often answer non-English questions better if they first translate the question into English and then reason. This trick, called self-translate, roughly doubles the cost of every query. A service that answers questions in many languages under a fixed compute budget therefore faces a real allocation problem. It can afford the extra translation pass for some queries but not all, and the choice decides which language communities get better answers. Treated as pure throughput maximization, this problem has a known failure mode from the fair-allocation literature: budgets drift toward the cheapest gains and starve the "hard-to-serve" groups.

This note argues that there is a second, less obvious failure mode, and that it comes from translationese. Translationese means the telltale fingerprints translated text carries: source-language word order, flattened vocabulary, and English-shaped phrasing. Most evidence that the English pivot helps comes from benchmarks that were themselves translated from English. On such items, translating back to English is close to undoing an earlier translation. On questions written natively, with local knowledge and native syntax, the pivot may gain little or even remove the cultural context the answer depends on. We propose two things. The first is to measure how translationese-like each incoming query is, using a cheap scorer. The second is to use that score to predict, per query, whether the costly English pass is worth buying, and to allocate a fixed budget of such passes across languages under an explicit fairness objective. We would test the whole approach on paired translated and native-origin benchmarks, because the translated benchmarks alone would hide the effect we care about.

Core Claim

The per-query benefit of English pivoting is not a fixed property of a language. It depends strongly on how translationese-like the query is. If that holds, two consequences follow. First, a scorer of query translationese gives a useful pre-inference signal for compute allocation, one that a language label or output confidence cannot fully replace. Second, efficiency–fairness frontiers for multilingual inference that are measured only on translated benchmarks are biased: they will overstate how much the pivot helps and misstate which languages a fair policy should prioritize. The new element is not welfare-based allocation, and it is not translationese measurement. It is the claim that one predicts the other, together with the demonstration that the fairness conclusion changes depending on whether evaluation data is native or translated.

Grounding in Prior Work

Several papers motivate the pivot side. Do Multilingual Language Models Think Better in English? introduces self-translate and reports that LLaMA-30B's average accuracy over XCOPA, XStoryCloze, XNLI, PAWS-X and MGSM rises from 43.1% to 46.6%. All five of those tasks are translated benchmarks, and the authors themselves note that translationese may favor the pivot. Two mechanistic papers point the same way. Can you map it to English? shows that patching English middle-layer activations into failing non-English runs repairs predictions. Multilingual Routing in Mixture-of-Experts steers middle-layer experts toward English-preferred ones and gains 1–2 points on MGSM and Global-MMLU. Again, all of this is measured on translated or parallel data. None of these papers asks whether the gain survives on native-origin questions, and none asks how to ration the cost.

A second group of papers constrains the design, because it shows translated evaluation can mislead. Is It Good Data for Multilingual Instruction Tuning or Just Bad Multilingual Evaluation? finds that native and translated training data look equivalent on MT-MMLU but separate clearly on CMMLU, C-Eval and generative XQuAD, where scores are 36.5% versus 28.5% Exact Match for Qwen-14B. ArabicMMLU reports GPT-4 falling from about 80% on translated MMLU to 72.5% on native Arabic exams. TUMLU and the Thunder Korean suite document native structures and cultural content that translation erases. Cross-Lingual Fairness Drift in LLM Moral Reasoning shows that cultural grounding collapses by up to 88% under translation. Taken together, these papers say that any allocation policy tuned on translated data must be validated on native data before it can be trusted.

Two lines of work supply the measurement tool. Translationese-index scores text by the log-likelihood ratio between two small (0.5B) language models, one trained on literal translations and one on idiomatic ones. It reports 95.8% in-domain classification accuracy, above 90% across genres, and 84.6% agreement with unanimous human pairwise judgments. An Investigation of Translationese in the Generations of Multilingual LLMs offers a cheaper stylistic classifier, a support vector machine over function words and part-of-speech patterns. Prior work uses both tools on outputs. We apply them to inputs, as a routing feature.

The objective comes from the welfare literature. Online Social Welfare Function-based Resource Allocation and Keep Everyone Happy allocate scarce resources under concave welfare functions that slide from total-utility maximization to protecting the worst-off. Planning to Fairly Allocate and Subsidy Allocations in the Presence of Income Shocks show the starvation pattern that pure throughput objectives produce. Those papers assume abstract, directly observed utilities. In our setting the utilities are latent and must be predicted from text, and that prediction is exactly where translationese enters. We deliberately do not import the online bandit machinery. A batch allocation over predicted gains is simpler and sufficient to test the claim; an online version is a follow-up only if the offline result holds.

Research Program

For each query $q$ in language $\ell$, let $g(q)$ be the predicted accuracy gain from running self-translate instead of direct inference. We estimate $g$ with a small calibrated model, such as logistic regression or a gradient-boosted tree, fit on a held-out split. It uses three groups of features: the query's translationese score, direct-inference confidence (answer entropy from the first pass, which is free), and the language identity. Given a budget $B$ of extra passes per batch, we choose a set $S$ of queries to pivot by maximizing an $\alpha$-fair welfare over per-language expected accuracy:

$$\max_{S,\ |S|\le B}\ \sum_{\ell} \frac{A_\ell(S)^{1-\alpha}}{1-\alpha},$$

Here $A_\ell(S)$ is the predicted mean accuracy in language $\ell$ when the queries in $S$ are pivoted. Setting $\alpha=0$ recovers throughput maximization, $\alpha\to1$ gives Nash welfare, and large $\alpha$ approaches max-min. Because the objective is concave in each $A_\ell$, a greedy marginal-gain procedure solves it well.

Translationese scorers exist off the shelf only for some languages, so they need extension. T-index was built on English and Chinese corpora, and the stylistic classifier covers German, Spanish, Greek and Pashto. For Arabic, Korean and Turkish we would train T-index pairs using the published recipe on native corpora paired with machine-translated text. We would validate each pair by checking that it separates the original-language and translated portions of the benchmarks we already hold. The scorer runs two 0.5B forward passes over the prompt only, which is small next to a second pass of a 7–30B model. We count this cost honestly in every compute budget.

The test models are open-weight LLaMA- and Qwen-family models at two sizes, matching the self-translate and routing papers. The translated evaluation side is XCOPA, XStoryCloze, XNLI, PAWS-X, MGSM, Belebele and MT-MMLU. The native side is CMMLU and C-Eval (Chinese), ArabicMMLU, TUMLU, the natively constructed portions of the Thunder Korean suite, and TyDi QA. Pairing by language lets us compare the same model, language and policy on translated versus native items.

Baselines

The naive reference points are uniform direct inference, which costs nothing extra, and uniform self-translate, which doubles the cost of every query. All other policies are reported as curves between these two endpoints, with total compute normalized identically.

The strongest existing method is self-translate from Do Multilingual Language Models Think Better in English?. It reports a five-task average of 46.6% versus 43.1% for direct inference with LLaMA-30B on translated benchmarks. We found no reported result on Lacuna for self-translate on native-origin benchmarks such as CMMLU, ArabicMMLU or TUMLU, and producing that number is itself a contribution. We also run external translate-test with NLLB-3.3B, as that paper did, and the middle-layer routing intervention of Multilingual Routing in Mixture-of-Experts on a mixture-of-experts model. That intervention moves Qwen3-30B on MGSM from 76.4% to 78.0% average non-English accuracy. It is a near-zero-cost competitor, so it sits at the cheap end of the frontier.

The decisive ablations remove the key ingredient while keeping everything else fixed. The first is a confidence-only router that pivots the lowest-confidence queries. The second is a language-level static policy that pivots a fixed fraction of each language's queries, chosen on the same tuning split. The third is our full predictor minus the translationese feature. Every router gets the same training split, the same hyperparameter search budget, and the same welfare objective. The proposal beats these baselines only if adding the translationese feature raises worst-language accuracy, or Nash welfare, at matched compute on native benchmarks by more than the across-seed and bootstrap confidence interval. A gain that appears only on translated benchmarks does not count.

Why not just…?

Why not just pivot whenever the model is unsure? This is the most serious threat, because output entropy is free and probably correlates with pivot benefit. The question is whether translationese carries information that confidence lacks. The way it could is this: a model can be confidently wrong on a native, culture-bound item where pivoting would hurt, and unsure on a translationese item where pivoting would help. The confidence-only ablation settles this directly. If adding the translationese score gives no gain beyond confidence on native data, we will report that the simple router suffices and that translationese matters only for evaluation, not for routing.

Why not just decide per language, pivoting only for weak languages? The welfare literature shows that the starvation problem appears with heterogeneous units, and languages are coarse units. If pivot benefit is roughly constant within a language, a per-language rule captures it and per-query scoring adds nothing. Native benchmarks mix translationese-like and native-like items within a single language, which is precisely what the T-index's graded scores are designed to expose. The comparison against the per-language rule therefore tests whether within-language variation in pivot benefit is real.

Why not just always translate with an external engine, or tune on translated data? External translate-test adds a dependency and its own errors, and Do Multilingual Language Models Think Better in English? shows translation quality becoming the bottleneck for low-resource languages. Tuning on translated data is the circular trap documented in Is It Good Data…?: it looks fine on translated tests and fails on native ones. Both strategies are included at their full-compute settings. If either dominates our curve on native data, that is a negative result we will report.

Decisive Experiments

The first check is item-level. For each benchmark, we regress the realized pivot gain (self-translate correct minus direct correct) on the input's translationese score while controlling for confidence. The core claim predicts a positive slope, and it predicts that native benchmarks have both lower average scores and smaller or negative average gains than their translated counterparts in the same language. If the slope is flat, the routing idea fails and we stop there.

The second check is the frontier comparison. We plot per-language worst-case accuracy and Nash welfare against extra compute for every policy, once on translated and once on native suites. The evaluation-bias claim predicts two things. One is that the ranking of policies, or the language each fair policy chooses to prioritize, differs between the two suites. The other is that frontiers tuned on translated data and deployed on native data underperform frontiers tuned on native data.

The third check concerns the fairness lens itself. We compare $\alpha=0$ with $\alpha$ near 1 and with large $\alpha$ at matched budget, and record which languages each setting starves. This makes concrete whether the throughput-maximizing allocation reproduces the bimodal-attention pattern described in Planning to Fairly Allocate in a multilingual service, and what accuracy cost removing that pattern carries.