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

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

cs.CLOct 7, 2026

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

Authors: HyeonSeok Lim, SeungWoo Song, Inho Won, Hoyun Song, Jihyo Kim, KyungTae Lim

Organizations: ETRI · KAIST · Department of Artificial Intelligence, Dankook University

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 https://github.com/lhsstn/LASEF.

Figures & tables

Figure 1: Overview of the Skeleton-Guided Reasoning Framework. (A) The process begins with an input query

q^{(\ell_{q})}

. (B) A structured skeleton

s^{(\ell_{s})}

is generated to abstract the problem logic. (C) The final reasoning and answer

\hat{a}

are produced in the answer language

\ell_{a}

, conditioned on the query and skeleton. (D) We define the language configuration space as

(\ell_{q},\ell_{s},\ell_{a})

and compare four primary strategies: standard CoT and skeleton-guided approaches, applied in both the target language (

\ell_{t}

) and English (

en

). (E) Effective configurations are identified through training-free benchmark analysis.

Method	Language	MGSM						MATH-500						PolyMath
Method	( $\ell_{q},\ell_{s},\ell_{a}$ )	zh	es	ko	th	sw	AVG.	zh	es	ko	th	sw	AVG.	zh	es	ko	th	sw	AVG.
Qwen2.5-7B-Instruct
CoT- $\ell_{t}$	$\ell_{t},-,\ell_{t}$	79.9	80.7	69.7	80.2	16.0	65.3	56.4	57.5	47.8	49.8	14.2	45.1	30.8	31.0	26.4	26.1	5.1	23.9
+ Skeleton	$\ell_{t},en,\ell_{t}$	80.7	83.8	71.0	79.0	22.9	67.5	55.4	57.7	49.3	51.9	20.4	46.9	30.6	32.1	28.0	27.6	7.4	25.1
CoT- $en^{\dagger}$	$en,-,\ell_{t}$	82.3	84.3	74.5	78.1	63.1	76.5	56.0	57.4	51.6	49.5	24.4	47.8	29.7	32.1	27.5	29.8	20.3	27.9
+ Skeleton	$en,en,\ell_{t}$	82.3	85.9	73.2	78.9	71.7	78.4	57.4	58.2	52.3	53.8	29.8	50.3	32.0	34.4	27.7	30.4	19.9	28.9

Table 1: Results on MGSM, MATH-500, and PolyMath.

(\ell_{q},\ell_{s},\ell_{a})

denote the query, skeleton, and answer languages, respectively. CoT-

en^{\dagger}

indicates queries translated into English using Google Translate, and the skeleton language is fixed to English unless otherwise noted. ‘AVG.’ denotes the average over five languages (zh, es, ko, th, sw). Greedy decoding throughout.

Language	$\ell_{t},en,\ell_{t}$		$en^{\dagger},en,\ell_{t}$
Language	CoT- $l_{t}$	+ Skeleton	CoT- $en^{\dagger}$	+ Skeleton
Qwen2.5-7B-Instruct
kk	40.08	51.42 ( +11.34 )	62.82	72.22 ( +9.40 )
ky	33.33	37.86 ( +4.53 )	61.92	65.69 ( +3.77 )
mn	27.16	33.33 ( +6.17 )	66.96	64.78 ( -2.18 )
ug	27.13	37.65 ( +10.52 )	57.14	60.37 ( +3.23 )

Table 2: Performance comparison on 19 low-resource languages (MGSM) using Qwen2.5-7B with English Skeleton. CoT-

en^{\dagger}

: English-translated queries via Google Translate. Parentheses show

\Delta

gains.

Panel of Figure 2 — Figure 2: Performance differences (Non-English $-$ English) on MGSM under CoT- $\ell_{t}$ with Qwen2.5-7B . The English skeleton is replaced with five Non-English skeletons (ZH, ES, RU, KO, TH); the same skeleton set is used in both panels. (a) Greedy decoding (single rollout). (b) Multi-rollout sampling ( $T{=}0.7$ , $N{=}5$ ). The corresponding English-skeleton reference values (English skeleton vs. no skeleton per language) are reported in Tab. 2 .

Figure 3: Translation ablation results. Each point is a (target, skeleton) pair, with

\Delta

from LLM-generated skeletons on the x-axis and translated skeletons on the y-axis. Points near the diagonal indicate effects robust to skeleton quality.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Table 3 — Table 3: The prompt instructs the model to produce a concise, high-level outline of the reasoning process, while omitting detailed computations and final answers.

Region	Languages	Script
Africa	Amharic	Ge’ez
Europe	Basque, Maltese, Armenian	Latin, Armenian
Central Asia	Kazakh, Kyrgyz, Uyghur, Pashto	Cyrillic, Arabic
South Asia	Nepali, Sinhala, Sindhi, Punjabi, Gujarati, Marathi, Odia, Tamil, Malayalam	Indic, Arabic
SE Asia & Pacific	Lao, Mongolian	Lao, Cyrillic

Appendix

Table 4: List of the 19 low-resource target languages included in the MGSM-LowResource dataset.

Target Lang	Acc (%)	$\Delta$	Target Lang	Acc (%)	$\Delta$
English (Baseline)	90.00	–	Quechua (qu)	89.60	-0.40
Tamil (ta)	93.60	+3.60	Guarani (gn)	90.00	0.00
Gujarati (gu)	91.60	+1.60	Basque (eu)	90.00	0.00
Kannada (kn)	91.20	+1.20	Amharic (am)	90.00	0.00
Khmer (km)	91.20	+1.20	Javanese (jv)	90.00	0.00
Sinhala (si)	91.20	+1.20	Uyghur (ug)	89.60	-0.40

Appendix

Table 5: Reasoning accuracy of Qwen2.5-7B on MGSM tasks using Round-trip Translation (English

\rightarrow

Pivot

\rightarrow

English). We used the target languages as pivots to validate the semantic consistency of the translated dataset.

\Delta

denotes the performance difference relative to the English baseline.

Statistic	Value
Total Target Languages	34
Samples per Language	250
Total Samples	8,500
Gemini Selection Ratio	62.1%
GPT Selection Ratio	37.9%
Avg. BT Similarity Score	0.96

Appendix

Table 6: Final statistics of the MGSM-LowResource dataset.

Code	Language	Family	Morph.	Order
Target languages
kk	Kazakh	Turkic	Agg	SOV
ky	Kyrgyz	Turkic	Agg	SOV
mn	Mongolian	Mongolic	Agg	SOV
ug	Uyghur	Turkic	Agg	SOV
hy	Armenian	IE (Armenian)	Agg †	SOV

Appendix

Table 7: Typological features of the 19 target and six skeleton languages. Morph. : dominant morphological type (Ana: analytic/isolating; Agg: agglutinative; Fus: fusional). Order : basic constituent order. Scripts are listed in Tab. 4 . † Armenian exhibits agglutinative nominal morphology. These features derive surface affinity and structural alignment as defined above.

Figure 4: Average token length of skeletons across query languages.

Figure 5: Average generated token counts of the Qwen-series models.

Figure 6: Quality evaluation results for five models in the PolyMath

Figure 7: Skeleton quality heatmap across language pairs. English skeletons yield highest quality (7.57), while Thai performs worst (5.65).

Language	Method	Problem	Conceptual	Reasoning	Calculation	Output Gen.
kk	CoT- $\ell_{t}$	100	61	129	33	116
kk	+SKELETON	91	54	105	18	135
ug	CoT- $\ell_{t}$	106	74	156	57	95
ug	+SKELETON	100	60	134	33	135
mn	CoT- $\ell_{t}$	107	78	157	47	88
mn	+SKELETON	108	62	142	41	115

Appendix

Table 8: Comparison of CoT-

\ell_{t}

and +SKELETON Results Across Different Reasoning Stages

Method	MGSM							MATH-500							PolyMath
Method	zh	es	ko	th	sw	te	AVG.	zh	es	ko	th	sw	te	AVG.	zh	es	ko	th	sw	te	AVG.
Llama-3.1-8B-Instruct
CoT $(\ell_{q}=\text{target})$	71.5	74.7	58.6	65.5	61.7	56.2	64.7	33.2	37.7	29.8	31.0	31.9	25.1	31.5	23.6	24.9	26.2	19.3	27.6	16.2	23.0
+ Skeleton	68.3	75.5	61.0	65.1	63.8	57.4	65.2	34.7	35.6	34.0	32.0	34.2	25.9	32.7	24.1	26.3	24.4	21.1	29.1	16.8	23.6
CoT $(\ell_{q}=\text{EN}^{\dagger})$	69.3	78.3	64.9	69.1	68.3	63.9	69.0	33.8	41.6	29.7	33.3	32.5	26.3	32.9	22.4	27.6	24.4	21.6	32.4	17.3	24.3
+ Skeleton	69.8	73.9	63.7	70.3	66.7	62.7	67.9	37.7	37.3	32.9	31.6	38.8	29.6	34.7	26.3	28.0	27.3	22.4	33.9	17.3	25.9

Appendix

Table 9: Multilingual mathematical reasoning results for the Llama 3.1 series (8B, 70B). The experiments use the MGSM, MATH-500, and PolyMath benchmarks.

\ell_{q}=\text{target}

indicates that questions are presented in the target language, while

\ell_{q}=\text{EN}^{\dagger}

denotes the use of English-translated questions. The skeleton language is fixed to English in all cases (

\ell_{s}=\text{EN}

). ‘AVG.’ denotes the average over the six languages.

Model	MGSM	MATH-500	PolyMath	AVG.	AVG. ∗
Ministral-3B	$+1.34$	$+1.79$	$+0.74$	$+1.29$	$+0.38$
Ministral-14B	$-1.10$	$+0.35$	$+1.08$	$+0.11$	$-0.12$

Appendix

Table 10: English-skeleton effect

\Delta

(pp) for the Ministral family under the CoT-

\ell_{t}

setting, averaged over the five languages of Tab. 1 (zh, es, ko, th, sw) per benchmark. ‘AVG.’ denotes the average over all five languages; AVG. ∗ excludes Swahili (sw), which retains relatively few valid samples.

Model	CoT- $\ell_{t}$		CoT- $en^{\dagger}$
Model	Base	+ Skeleton	Base	+ Skeleton
Qwen2.5-7B	4.75	4.91 ( +0.16 )	5.12	5.82 ( +0.70 )
Qwen2.5-14B	6.62	8.40 ( +1.78 )	7.13	7.56 ( +0.43 )
Qwen2.5-72B	9.27	9.97 ( +0.70 )	8.92	8.96 ( +0.04 )

Appendix

Table 11: Skeleton gains on the two hardest PolyMath subsets (Top + High; 250 of 500 problems), as Exact-Match accuracy (%) averaged over five languages (zh, es, ko, th, sw). Parentheses show the gain (

\Delta

) from adding the English skeleton. Compared with the full-set gains (Tab. 1 ), improvements shrink on the harder subset, consistent with the diminishing effect of skeletons as difficulty increases.

Target	Skeleton	$\Delta$ (%)	$p$
Positive effects (Non-English skeleton $>$ English skeleton)
mt	ru	$+8.62$	0.001
mt	th	$+6.90$	0.008
mt	es	$+6.61$	0.041
Negative effects (English skeleton $>$ Non-English skeleton)
ug	th	$-10.25$	0.002

Appendix

Table 12: Pairs that reached nominal significance in McNemar’s test before correction for multiple comparisons (

p<0.05

\Delta

denotes the accuracy difference between the non-English skeleton and the English skeleton.

Target	Skel.	MGSM $\Delta$	MSVAMP $\Delta$	Outcome
hy	ko	$+5.17$	$+2.75$	Retained
lo	zh	$+3.63$	$+2.63$	Retained
lo	th	$+2.81$	$+2.42$	Retained
eu	ko	$+6.32$	$0.00$	Disappeared
mt	es	$+6.61$	$+2.00$	Attenuated
mt	th	$+6.90$	$+0.43$	Attenuated

Appendix

Table 13: Cross-benchmark transfer of skeleton-language effects from MGSM to MSVAMP ( Qwen2.5-7B , CoT-

\ell_{t}

\Delta

: accuracy difference (pp) between the non-English and the English skeleton. Pairs were fixed using only the MGSM results before inspecting MSVAMP. Across all 35 target–skeleton cells, the sign is preserved in 23/35 (66%).

Method	MATH-500 (te)	PolyMath (te)	MGSM (te)
Qwen2.5-7B-Instruct
CoT- $\ell_{t}$	29.8	12.2	30.40
+SKELETON	27.8	12.4	33.60
CoT- $en^{\dagger}$	35.7	17.3	60.40
+SKELETON	33.9	17.7	54.00
Qwen2.5-14B-Instruct

Appendix

Table 14: Telugu (te) Accuracy (%) on MATH-500, PolyMath, and MGSM

Figure 8: Performance differences (Non-English − English) under the CoT- $\ell_{t}$ setting using Qwen2.5-7B across all 34 target languages. While English generally serves as a stable default skeleton language, several target languages exhibit gains with specific Non-English skeletons, indicating that skeleton effectiveness is language-dependent.

Explore similar work

LANG: Reinforcement Learning for Multilingual Reasoning with Language-Adaptive Hint Guidance
May 21, 2026Yuchun Fan, Bei Li, Peiguang Li +9LLM Reasoning Strategies Multilingual Language Models
NLP Lab, School of Computer Science and Engineering, Northeastern University, Shenyang, China · Meituan Inc. 3NiuTrans Research, Shenyang, China
Beyond Input Understanding: Diagnosing Multilingual Mathematical Reasoning with Directed Acyclic Trace Graphs
May 26, 2026Jiaqiao Zhang, Zhoujun Li, Raoyuan Zhao +5Mathematical Reasoning Reasoning Traces
Southwest University · LMU Munich & MCML
x1: Learning to Think Adaptively Across Languages and Cultures
Apr 18, 2026Yangfan Ye, Xiaocheng Feng, Xiachong Feng +8LLM Reasoning Strategies
Harbin Institute of Technology · Peng Cheng Laboratory · The University of Hong Kong +1

May 21, 2026, cs.CLJ/K move · Enter open · S save

LANG: Reinforcement Learning for Multilingual Reasoning with Language-Adaptive Hint Guidance

Yuchun Fan, Bei Li, Peiguang Li, Yilin Wang+8

NLP Lab, School of Computer Science and Engineering, Northeastern University, Shenyang, China · Meituan Inc. 3NiuTrans Research, Shenyang, China

LLM Reasoning Strategies Multilingual Language Models Language Pairs Goal-Conditioned Dynamic Manipulation Guidance

Open paper AskPDF arXiv

Reinforcement learning has proven effective for enhancing multi-step reasoning in large language models (LLMs), yet its benefits have not fully translated to multilingual contexts. Existing methods struggle with a fundamental trade-off: prioritizing input-language consistency severely hampers reasoning quality, while prioritizing reasoning often leads to unintended language drift toward English. We address this challenge with LANG, a novel framework that leverages language-conditioned hints to guide exploration in non-English reasoning tasks. Our method incorporates two key mechanisms to prevent dependency on these hints: a progressive decay schedule that gradually withdraws scaffolding, and a language-adaptive switch that tailors learning horizons to specific language difficulties. Empirical results on challenging multilingual mathematical benchmarks reveal that LANG substantially enhances reasoning performance without compromising language consistency. Moreover, we show that our framework generalizes beyond mathematics, fostering more consistent language alignment across model layers

Abstract

Explore similar work

LANG: Reinforcement Learning for Multilingual Reasoning with Language-Adaptive Hint Guidance

Beyond Input Understanding: Diagnosing Multilingual Mathematical Reasoning with Directed Acyclic Trace Graphs

x1: Learning to Think Adaptively Across Languages and Cultures