Which Language Should a Skeleton Speak? Language Choices in Multilingual Reasoning

Lim, HyeonSeok; Song, SeungWoo; Won, Inho; Song, Hoyun; Kim, Jihyo; Lim, KyungTae

doi:10.18653/2026.findings-emnlp.1500

How the language of a reasoning skeleton affects multilingual mathematical reasoning in LLMs

Findings of EMNLP 2026

HyeonSeok Lim¹, SeungWoo Song², Inho Won², Hoyun Song^2,3, Jihyo Kim², KyungTae Lim^2,†

¹ETRI ²KAIST ³Department of Artificial Intelligence, Dankook University

^†Corresponding author

Abstract

Skeleton-based reasoning prompting is a promising training-free approach for structuring LLM reasoning, but prior work largely assumes an English-centric setting. We propose the Language-Aware Skeleton Exploration Framework (LASEF) to study skeleton-language choice in multilingual mathematical reasoning. Across math benchmarks, model scales, and languages, we show that English skeletons yield a small positive tendency on average, most visible for smaller models and low-resource languages. However, few language-level gains remain significant after correction, and English is not universally optimal. Combining greedy decoding, multi-rollout evaluation, translation ablation, and cross-benchmark validation, we further find three patterns of skeleton-language effects: directionally consistent, evaluation- and benchmark-dependent, and asymmetric negative. These effects cannot be fully explained by generation quality alone. Overall, skeleton language is a context-dependent design variable that requires multi-level exploration. All resources are released at github.com/lhsstn/LASEF.

Keywords Skeleton-based reasoning Multilingual reasoning Multilingual math reasoning Chain-of-Thought (CoT) English-pivot reasoning Cross-lingual transfer Low-resource languages Training-free prompting LASEF MGSM

Method

A skeleton is a short, abstract description of a problem's structure (what is known, what is unknown, which concepts or tools apply) that guides the model's reasoning without giving solution steps or the answer. In a multilingual setting, skeleton-based reasoning involves three language choices: the query language $\ell_q$ , the skeleton language $\ell_s$ , and the answer language $\ell_a$ used for both reasoning and the final answer. LASEF is a training-free analysis framework (Figure 1) that varies only these languages while keeping the skeleton format fixed:

Skeleton generation. From the query $q^{(\ell_q)}$ , the model generates a skeleton $s^{(\ell_s)}$ under one rule: describe what the problem structure is, never how to solve it. Skeletons average about 120 tokens regardless of language.
Skeleton-guided reasoning. The model then produces the answer conditioned on the query and skeleton, $\hat{a} = f\big(q^{(\ell_q)},\, s^{(\ell_s)},\, \ell_a\big),$ where $f$ is the pre-trained LLM.
Language configurations. For a target language $\ell_t$ , four baselines are compared: target-language CoT $(\ell_t, -, \ell_t)$ , the same with an English skeleton $(\ell_t, en, \ell_t)$ , English-pivot CoT on a Google-translated query $(en, -, \ell_t)$ , and the same with an English skeleton $(en, en, \ell_t)$ . The study then replaces the English skeleton with Chinese, Spanish, Russian, Korean, and Thai skeletons.
Multi-level evaluation. Configurations are scored with exact match on samples whose answer is in the target language, under greedy decoding and multi-rollout sampling (T = 0.7, N = 5), with a translation ablation (English skeletons translated by GPT-5-mini, so content is fixed and only surface language varies) and cross-benchmark checks on MSVAMP.

Experiments use Qwen2.5-7B/14B/72B-Instruct (Llama 3.1 results are in the paper's appendix) on MGSM (elementary arithmetic), MATH-500 (intermediate), and PolyMath (advanced), with target languages ranging from high-resource (zh, es) to low-resource (19 low-resource MGSM languages such as Kazakh, Uyghur, Basque, Maltese, and Lao).

Main Results

English skeletons help a little on average. On 19 low-resource MGSM languages with Qwen2.5-7B, English skeletons improve 13 of 19 languages, with mean gains of +2.38 points (33.99 → 36.37, target-language query) and +2.02 points (58.06 → 60.08, English-translated query). After FDR correction, language-level gains are significant only for Kazakh (+11.34) and Uyghur (+10.52); Maltese drops by 5.65.
Translation matters more than the skeleton. Translating the query into English raises the low-resource average from 33.99 to 58.06 (+24.07), versus +2.02 from adding English skeletons on top.
Gains shrink with model size and difficulty. For Qwen2.5-7B, English skeletons add +2.2 on MGSM, +1.8 on MATH-500, and +1.2 on PolyMath; Swahili on MGSM (translated query) rises from 63.1 to 71.7. Qwen2.5-72B instead drops by 0.2 on MGSM (target-language query) and on PolyMath (translated query).
English is a stable but not universal anchor. Replacing the English skeleton with zh, es, ru, ko, or th lowers the average by 0.63 to 2.60 points, yet some pairs gain under both greedy and multi-rollout evaluation: Basque with a Korean skeleton (+6.32 → +2.78), Armenian–Korean (+5.17 → +4.61), and Lao–Thai (+2.81 → +3.18).
Three patterns, not explained by skeleton quality alone. Besides these directionally consistent pairs, Maltese shows evaluation-dependent gains (Russian skeleton +8.62 under greedy decoding, attenuated or reversed under multi-rollout and on MSVAMP), and Mongolian–Russian (−5.00) and Uyghur–Chinese (−6.45) show asymmetric negative transfer despite shared script or geography. With translated skeletons the effects keep their sign (eu–ko +4.42, mt–ru +5.15, mn–ru −3.43, ug–zh −6.94).

**Table 1.** Exact-match accuracy (%) with and without English skeletons, greedy decoding. (ℓ_q, ℓ_s, ℓ_a) are the query, skeleton, and answer languages; en^† is a query translated into English with Google Translate. AVG. is the mean over zh, es, ko, th, and sw. Subset: Swahili (sw) and AVG. columns only; per-language results for zh, es, ko, and th are in the full table in the paper. Skeleton rows (LASEF with an English skeleton) are highlighted. (Table 1 in the paper.)
Model	Method	(ℓ_q, ℓ_s, ℓ_a)	MGSM		MATH-500		PolyMath
Model	Method	(ℓ_q, ℓ_s, ℓ_a)	sw	AVG.	sw	AVG.	sw	AVG.
Qwen2.5-7B-Instruct	CoT-ℓ_t	ℓ_t, −, ℓ_t	16.0	65.3	14.2	45.1	5.1	23.9
	+ Skeleton	ℓ_t, en, ℓ_t	22.9	67.5	20.4	46.9	7.4	25.1
	CoT-en^†	en, −, ℓ_t	63.1	76.5	24.4	47.8	20.3	27.9
	+ Skeleton	en, en, ℓ_t	71.7	78.4	29.8	50.3	19.9	28.9
Qwen2.5-14B-Instruct	CoT-ℓ_t	ℓ_t, −, ℓ_t	45.0	77.9	30.1	52.6	14.7	29.1
	+ Skeleton	ℓ_t, en, ℓ_t	50.7	79.1	35.2	55.1	15.0	30.3
	CoT-en^†	en, −, ℓ_t	74.5	82.4	39.1	53.4	24.0	31.0
	+ Skeleton	en, en, ℓ_t	74.5	83.3	40.4	54.6	26.9	32.0
Qwen2.5-72B-Instruct	CoT-ℓ_t	ℓ_t, −, ℓ_t	68.0	84.5	46.3	59.6	31.0	34.8
	+ Skeleton	ℓ_t, en, ℓ_t	70.1	84.3	54.3	61.9	32.0	35.7
	CoT-en^†	en, −, ℓ_t	85.5	86.3	56.3	61.5	40.6	36.3
	+ Skeleton	en, en, ℓ_t	86.7	86.8	60.0	62.4	39.3	36.1

Two heatmaps of accuracy differences between non-English and English skeletons for 19 low-resource target languages (rows) and five skeleton languages ZH, ES, RU, KO, TH (columns), under greedy decoding (left) and five-rollout sampling (right). Most cells are negative (blue); positive (red) cells include Maltese with Russian, Thai, and Spanish skeletons under greedy decoding, and Basque and Armenian with Korean and Lao with Thai under both settings. Average rows range from -0.63 to -2.36 (greedy) and -1.27 to -2.60 (multi-rollout).

Figure 2. Accuracy difference (non-English skeleton − English skeleton, points) on MGSM with Qwen2.5-7B and target-language queries, for skeletons in ZH, ES, RU, KO, and TH. (a) Greedy decoding: averages are −0.67, −1.64, −0.63, −1.20, and −2.36. (b) Multi-rollout sampling (T = 0.7, N = 5): averages are −1.35, −1.27, −1.45, −1.81, and −2.60. Largest positive cells under greedy decoding: MT–RU +8.62, MT–TH +6.90, MT–ES +6.61, EU–KO +6.32, HY–KO +5.17; largest negative: UG–TH −10.25. (Figure 2 in the paper.)

@inproceedings{lim-etal-2026-language, title = "Which Language Should a Skeleton Speak? Language Choices in Multilingual Reasoning", author = "Lim, HyeonSeok and Song, SeungWoo and Won, Inho and Song, Hoyun and Kim, Jihyo and Lim, KyungTae", editor = "Lee, Hung-Yi and Dev, Sunipa and Li, Jessy and Elliott, Desmond", booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026", month = oct, year = "2026", address = "Budapest, Hungary", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2026.findings-emnlp.1500/", doi = "10.18653/2026.findings-emnlp.1500", pages = "31933--31957", ISBN = "979-8-89176-462-0" }

Which Language Should a Skeleton Speak? Language Choices in Multilingual Reasoning

Abstract

Method

Main Results

BibTeX