Choosing the Language to Think In: Uncertainty-Selected Distillation for Multilingual Students

Multilingual language models solve a math problem posed in Swahili much better if they reason about it in English first. They answer a question about local social norms better if they reason in the language where that norm is discussed. Current methods for teaching small models to reason across languages pick one reasoning language for everything, almost always English. We propose a different recipe. A large teacher model attempts each unlabeled training question several times in English and several times in the question's own language. We measure how much its final answers disagree within each language and across the two. The student is trained only on the teacher's most self-consistent reasoning, and only in the language where the teacher was most consistent. The result is a small student that writes a one-token language tag before it starts reasoning, so it decides per question whether to pivot through English. It does this without a translation pipeline at inference and without needing gold labels to build its training data.

This matters because the two halves of the evidence currently contradict each other. Forcing English reasoning helps math enormously and hurts culture and safety. Small distilled students also copy whatever errors are in the teacher's traces. One signal can address both problems: semantic uncertainty, meaning how much the teacher's answers vary in meaning when it is sampled repeatedly. It can serve as a label-free quality filter on the traces and as the rule that decides which language to reason in.

Core Claim

Our central claim is that, for a fixed teacher and an unlabeled pool of target-language questions, answer-level semantic entropy computed separately for English-pivot and native-language reasoning gives three things at once. It ranks traces by correctness. It identifies which reasoning language suits each question. And it produces a training set from which a compact student learns to route between languages per question. Concretely, the student should match an always-English student on multilingual math, beat it on culture-heavy benchmarks, and keep answering in the user's language.

The new element is not uncertainty filtering of distillation data, which Towards Label-Free Biological Reasoning Synthetic Dataset Creation via Uncertainty Filtering already did in English. That paper kept the 10% lowest-uncertainty traces under a metric mixing self-consistency and perplexity. What we add is treating the reasoning language as a variable chosen by uncertainty and distilling that choice into the student.

Where This Sits

Three lines of work motivate the routing half. Language models are multilingual chain-of-thought reasoners introduced the multilingual grade-school math benchmark MGSM and found English-pivot reasoning beats native reasoning at large scale. Language Matters showed that large reasoning models drift into an English or Chinese "hub" while thinking. The same paper showed that prefilling a single word after the thinking tag steers the thinking language with over 97% consistency. English steering raised MATH-500 accuracy by 26.8% on average for an 8B model, but it raised toxicity by 1–3.5% and lost ground on CulturalBench. Language Specific Knowledge went further: its LSKExtractor clusters queries offline, assigns each cluster an "expert language," and translates test queries into it, gaining up to 10.4% relative on CultureAtlas.

These papers establish the gap: no single reasoning language is best. They also constrain the design, because cluster-level assignment and external translation are exactly what we aim to replace with a per-query, internal choice. The prefilling result supplies the mechanism we reuse: a short language tag is enough to control the student's thinking language.

The distillation half builds on On-Policy Delta Distillation for Multilingual Math Reasoning. That paper distills Qwen3-30B into Qwen3-1.7B and 8B, using the gap between the teacher's post-trained and base distributions as the distillation target. It found that English-only training collapses Japanese response rates to 29.6%, while balanced multilingual data keeps them near 90%. It distills every token of every trace and studies only English, Korean, and Japanese. XCOT and PLUG hard-code English as the reasoning anchor. XCOT does so through code-switching and cross-lingual output matching; PLUG does so by writing a translated instruction and an English draft before every answer. Both distill or train on all traces equally. PLUG also pays for the pivot text at every inference step.

Language Imbalance Driven Rewarding assumes English-origin answers are always preferable, which is the very assumption the cultural results above undermine. SCOTT and Mitigating Misleading Chain-of-Thought Reasoning with Selective Filtering warn that students take shortcuts around rationales and that small models are sometimes hurt by reasoning at all. SelF-Reasoner handles this with a separately trained entailment checker. We instead use the teacher's own uncertainty, and in a multilingual setting none of them consider.

Research Program

The uncertainty signal is chosen to survive the multilingual setting. Token-level perplexity is a poor choice across scripts, because tokenizers split Swahili or Telugu into many more pieces than English. Perplexity would then measure tokenization rather than confidence, which is a calibration drift the uncertainty literature flags as unresolved. We therefore measure uncertainty on final answers only. For a question $q$ and reasoning language $r$ (English or the question's native language), we sample $K$ teacher traces that end in an answer in the target language. We group answers that mean the same thing: identical numbers for math, identical options for multiple choice. We then compute

$$H_r(q) = -\sum_{c} p_r(c \mid q)\log p_r(c \mid q),$$

where $c$ ranges over the answer groups and $p_r(c\mid q)$ is the fraction of the $K$ samples falling in group $c$. We also record whether the majority answers under the two reasoning languages agree. Disagreement means one route is losing information, either through translation loss or through knowledge that exists only in one language.

Each question gets the reasoning language $r^*(q) = \arg\min_r H_r(q)$. We keep the question only if $H_{r^*}(q)$ falls below a threshold set as a per-language quantile. The per-language quantile follows the per-class stratification that made the biological filtering work, and it prevents low-resource languages from being filtered out wholesale. One majority-answer trace in $r^*$ becomes the training target, prefixed with a language tag such as "Okay" for English or a native opener. The final answer is always written in the user's language, which guards against the language drift reported in the delta-distillation work.

The student is trained by standard supervised fine-tuning on these tagged traces. In a second stage it is trained on-policy with the delta-distillation objective, with the teacher conditioned on the same tag. This lets us check that the selection gains add to, rather than replace, the best existing distillation loss.

The teacher is Qwen3-30B-A3B, which Language Matters analyzed and the delta-distillation paper used as a teacher. The students are Qwen3-1.7B and Qwen3-8B. The unlabeled question pool comes from translated GSM8K-style math, as in the XCOT-INSTRUCT construction, plus cultural multiple-choice questions in the CultureAtlas and BLEnD style. Questions that overlap with any evaluation split are removed. We use $K=8$ samples per language. This sampling cost is paid once, offline, and it is the main compute price of the method.

Baselines

The naive reference point is the student before any distillation, prompted to reason natively and, separately, prompted with an English-pivot prefix. These two numbers show how much any training adds.

The strongest existing method is multilingual on-policy delta distillation from On-Policy Delta Distillation for Multilingual Math Reasoning with a balanced language split. It is the only cited method that distills reasoning into small students across languages while controlling drift. For Qwen3-1.7B it reports average scores rising from 40.9 to 51.9 in Korean and from 37.2 to 52.0 in Japanese, and an English–Korean Global-MGSM gap shrinking from 13.4 to 9.3 points. We found no reported result on Lacuna for this method on Swahili, Bengali, Telugu, or Thai, or on any cultural benchmark, so we will rerun it on our pool. It gets the same teacher, the same students, the same question pool before filtering, and the same number of optimizer steps and learning-rate sweep. Beating it means matching it within noise on MGSM and Global-MGSM while improving CultureAtlas and BLEnD accuracy and keeping the target-language response rate at or above its roughly 90%.

The always-English variant uses our identical pipeline with $r$ fixed to English. This is the XCOT and PLUG style of anchoring, but without paying for translation tokens at inference. The always-native variant fixes $r$ to the question's language. A distilled LSKExtractor baseline uses the cluster-level expert language from Language Specific Knowledge to pick each training trace's language, isolating per-query versus per-cluster routing.

Two ablations isolate the key ingredients. The first keeps the language choice but replaces uncertainty filtering with a random subset of the same size. The second keeps the filtering but uses single-language CoCoA-style uncertainty without the cross-language comparison.

Finally, a gold-correctness filter keeps only traces whose answers match the labels. It is an oracle for the filtering half and the real threat on math.

Why not just…?

Why not just filter traces by gold-answer correctness, since MGSM-style math has labels? On labeled math this is a serious competitor and may match us. Our claim is not that uncertainty beats labels. The claim is that uncertainty lets us use unlabeled target-language questions, which are far more plentiful in low-resource languages, and that it supplies the language choice, which correctness alone does not. The experiment settles this directly. We compare gold-filtered training on a small labeled set against uncertainty-selected training on a pool several times larger. We also report what fraction of uncertainty-selected traces are wrong when checked against labels, which directly measures confident-but-wrong leakage. If gold filtering on the small set wins, the label-free argument fails for math and we restrict the claim to the routing benefit.

Why not just always reason in English with multilingual answers? On math this is likely to tie us, given the 26.8% English-steering gain in Language Matters. On culture and safety, however, the same paper and LSKExtractor show English reasoning losing. The always-English variant run through our identical pipeline will show whether routing costs anything on math and gains anything on CultureAtlas, BLEnD, CulturalBench, and LMSYS-Toxic. If the routed student does not beat it on the cultural sets, the routing idea is refuted.

Why not just distill LSKExtractor's cluster-level language assignment? It is cheaper and needs no per-query sampling. Clusters, though, average over questions whose best language differs, and LSKExtractor needed labels to score each cluster. The distilled-LSKExtractor baseline uses the same student and data. Our per-query routing must beat it on the cultural benchmarks, where its 10.4% relative gain on CultureAtlas is the reference effect size.

Why not just sample many student answers at test time and take a majority vote? That multiplies inference cost by the number of samples. SelF-Reasoner also argues that small models produce correlated wrong rationales that voting reinforces. We report a student self-consistency baseline at matched inference compute as a separate axis, while our student runs one greedy pass.

Proposed Experiments

Evaluation covers MGSM across all eleven languages, with Swahili, Bengali, Telugu, and Thai reported separately as the low-resource subset, plus Global-MGSM and MATH-500 under translated prompts for math. CultureAtlas, BLEnD, and CulturalBench cover culture. LMSYS-Toxic covers safety, since the English-hub toxicity increase is a reported cost. For every system we report accuracy, the fraction of answers in the user's language, the distribution of chosen reasoning languages per benchmark, and generated tokens per answer, the last to compare against PLUG-style pivoting.

Four checks decide the proposal. First, on the labeled subset, lower $H_{r^*}$ must predict correctness in every language, not just English. We report per-language calibration curves, because this is where tokenization and script effects would break the signal. Second, the routed student must pick English mostly on math and the native language more often on cultural questions. If it routes uniformly, it has learned nothing. Third, the routed student must beat always-English on cultural accuracy while staying within noise of it on MGSM. Fourth, uncertainty-selected data must beat random selection of equal size by a margin comparable in kind to the biological result (0.57 versus 0.48 for random 10% and 0.52 for the full pool, in Towards Label-Free Biological Reasoning Synthetic Dataset Creation via Uncertainty Filtering), and the cross-language comparison must add over single-language uncertainty.

If the first check fails for low-resource languages, the method reduces to English-anchored distillation. That would itself be a useful negative result about multilingual uncertainty calibration.