Choosing the Right Language Mode at Inference Time for Multilingual Reliability

Ekata Mitra; Ameeta Agrawal

cs.CLSep 4, 2026

Choosing the Right Language Mode at Inference Time for Multilingual Reliability

Authors: Ekata Mitra, Ameeta Agrawal

Organizations: PortNLP, Portland State University, USA

Abstract

Multilingual large language models often struggle to reason in low- to mid-resource languages. Prior work has shown that translation can improve multilingual reasoning by helping models access stronger English-centric representations. This raises a central question: How much translation is needed for multilingual large language models to reason reliably, and when does more translation instead trigger interference and overconfidence? Using LLaMA and Qwen models, we run extensive experiments varying text scope and language mode (target-only, English-only, bilingual) to evaluate both accuracy and reliability. Our results reveal a clear trade-off: English context often improve understanding and recover errors caused by non-English comprehension, yet adding redundant bilingual context intensifies interference. We address this trade-off with Reliability-Aware Adaptive Inference (RAAI), a training-free test-time framework that (i) performs Expected Calibration Error (ECE)-aware routing and prompt fusion, and (ii) uses a mid-layer Risk Index (RI) to gate sequential reasoning, allocating compute only when it is likely to help and suppressing harmful bilingual redundancy. Across two model families, RAAI enhances accuracy by 25-37.7% on low-resource languages and lowers calibration error by approximately 3-6%, with the most pronounced benefits in the lowest-resource language tiers.

Figures & tables

Prior work	Language Mode	Mapped Condition
Target-language multilingual QA ( Bandarkar et al., 2025 ; Yang et al., 2023 )	T	Question, passage, and choices remain in the target dialect throughout, with no English support at inference (our T-mode baseline)
English-pivot / translate-test ( Shi et al., 2023 ; Qin et al., 2023 ; Zhu et al., 2024 )	EN	All input fields are translated into English and scored with the same decoding scheme as T, isolating the effect of full English pivoting (our EN-mode)
Bilingual prompting ( Shi et al., 2023 ; Zhu et al., 2024 ; Huang et al., 2023 )	TEN	Target and English fields are provided jointly, with translation scope varied independently across Q/P/QP/C/All (our TEN-mode)

Table 1: Mapping prior multilingual strategies onto our controlled T/EN/TEN taxonomy.

Panel of Figure 1 — Figure 1: An example from the Belebele dataset illustrating passage, question, and choices.

Figure 2: Correctness taxonomy by resource tier (H/M/L) across text scopes [Q/P/QP/C/All] for Belebele dataset using LLama model.

	LLaMA 3.1-8B Instruct						Qwen 3-4B
Resource	T		EN		TEN		T		EN		TEN
	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$
Belebele
High	81.26	9.15	81.79	10.05	81.78	8.75	80.28	9.14	87.73	7.43	88.69	8.25
Mid	68.88	14.05	65.15	12.42	61.79	9.81	66.28	10.67	86.55	8.64	79.86	10.92
Low	44.59	12.36	44.30	12.81	44.30	14.05	45.95	9.42	85.44	6.48	81.65	9.31

Table 2: Analysis resource-tier results for Belebele and MMLU-ProX-Lite (LLaMA/Qwen). Accuracy and ECE for High/Mid/Low tiers under T/EN/TEN; bold marks the best accuracy and lowest ECE per resource tier (row-wise).

Figure 3: Layer - wise comparison for Qwen - MMLUProXLite and LLaMA - Belebele under ‘ALL’ mode : (a) accuracy curves per dialect with a black mean; (b) ECE curves per dialect with a black mean and mean confidence overlaid (red dashed); (c) entropy heatmaps (dialect × layer). All metrics are aggregated over dialects from logit - lens probes, highlighting late - layer accuracy gains alongside calibration and uncertainty shifts.

	LLaMA 3.1-8B-Instruct								Qwen3-4B
Tier	T Static		Best Static		RouteGate		SeqGate		T Static		Best Static		RouteGate		SeqGate
	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$	Acc $\uparrow$	ECE $\downarrow$
Belebele
High	81.26	9.15	81.79	10.05	83.58	5.44	82.58	6.52	80.28	9.14	88.69	8.25	83.77	5.19	88.54	7.69
Mid	68.88	14.05	68.88	14.05	73.19	6.56	67.68	10.62	66.28	10.67	86.55	8.64	69.33	10.76	62.96	11.68
Low	44.59	12.36	44.59	12.36	55.19	7.71	69.70	11.34	45.95	9.42	85.44	6.48	71.14	13.19	83.64	8.35

Table 3: Comparison of T- and Best- Static baseline (from Table 2 , along with the proposed RAAI- RouteGate and RAAI- SeqGate by resource tier. Bold marks the best accuracy and lowest ECE per row.

	LLaMA 3.1-8B-Instruct					Qwen3-4B
Tier	$\rho(\mathrm{RI},\mathrm{Acc}\!\downarrow)$	$\rho(\mathrm{RI},\mathrm{ECE}\!\uparrow)$	N	MedErr	AUROC $\uparrow$	$\rho(\mathrm{RI},\mathrm{Acc}\!\downarrow)$	$\rho(\mathrm{RI},\mathrm{ECE}\!\uparrow)$	N	MedErr	AUROC $\uparrow$
Belebele
High	-0.54 (0.013)	+0.55 (0.012)	20	17.88	0.73	-0.29 (0.207)	-0.13 (0.583)	20	18.64	0.58
Mid	-0.21 (0.457)	-0.21 (0.457)	15	31.25	0.50	-0.55 (0.033)	+0.75 (0.001)	15	30.00	0.60
Low	-0.58 (0.275)	+0.68 (0.275)	10	55.41	0.72	-0.87 (0.001)	+0.74 (0.014)	10	54.05	1.00
MMLU-ProX-Lite

Table 4: RI validation by resource tier: Spearman

\rho

with Accuracy (Acc) and ECE, N, Median Error (MedErr), and AUROC (p-values in parentheses) for LLaMA and Qwen. Median Error and AUROC use within each resource tier as the positive-class threshold.

Example 1 (Bengali): T works, EN fails
(P)	Vision, or the ability to see depends on visual system sensory organs or eyes …
(Q)	What is the ability to see dependent upon?
(C)	A) Visual system sensory organs B) The requirements of an organism C) A sensitivity to certain wave-lengths D) Varying degrees of acuity
Ans.	A
Pred.	T = A (correct); EN = blank/invalid; TEN = A (correct).
Example 2 (Yoruba): T fails, EN fails, TEN works

Table 5: Qualitative analysis showing example instances. Examples shown in English.

Appendix figures & tables30 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 4 — Figure 4: Belebele example illustrating translation granularity (Q/P/QP/C/All).

Panel of Figure 5 — Figure 5: Evaluation prompt templates (abridged).

Panel of Figure 6 — Figure 6: Sequential bilingual reasoning prompt used for MMLU-ProX-Lite.

Resource	T				EN				TEN
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	81.26	76.58	9.15	0.64	81.79	74.75	10.05	0.70	81.78	76.34	8.75	0.66
Mid	68.88	60.01	14.05	1.00	65.15	63.11	12.42	0.93	61.79	64.67	9.81	0.90
Low	44.59	39.38	12.36	1.29	44.30	47.36	12.81	1.20	44.30	45.67	14.05	1.22

Appendix

Table 6: Static-mode full metrics by resource tier (Belebele, LLaMA). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				EN				TEN
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	34.65	38.78	11.02	1.77	31.02	40.50	13.19	1.74	38.84	33.43	12.16	1.92
Mid	25.64	33.70	13.13	1.91	35.19	34.13	10.44	1.91	31.57	27.27	11.67	2.05
Low	23.47	29.72	10.02	1.99	29.35	32.83	13.10	1.95	33.70	28.44	13.46	2.05

Appendix

Table 7: Static-mode full metrics by resource tier (MMLU-ProX-Lite, LLaMA). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				EN				TEN
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	80.28	82.77	9.14	0.55	87.73	85.78	7.43	0.50	88.69	85.20	8.25	0.48
Mid	66.28	63.51	10.67	0.95	86.55	87.87	8.64	0.43	79.86	86.81	10.92	0.40
Low	45.95	45.67	9.42	1.22	85.44	86.72	6.48	0.46	81.65	78.81	9.31	0.61

Appendix

Table 8: Static-mode full metrics by resource tier (Belebele, Qwen). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				EN				TEN
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	37.78	45.37	13.83	1.71	36.27	44.47	13.02	1.71	45.12	49.33	15.06	1.58
Mid	32.67	40.35	13.82	1.79	34.52	43.22	16.69	1.73	46.38	55.70	21.51	1.41
Low	28.57	38.15	12.81	1.83	35.87	44.28	18.24	1.72	44.57	55.30	17.63	1.43

Appendix

Table 9: Static-mode full metrics by resource tier (MMLU-ProX-Lite, Qwen). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				TEN				RouteGate
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	81.26	76.58	9.15	0.64	83.23	78.20	8.55	0.61	83.58	79.92	5.44	0.59
Mid	68.88	60.01	14.05	1.00	71.52	68.11	11.10	0.82	73.19	74.00	6.56	0.74
Low	44.59	39.38	12.36	1.29	55.22	52.58	10.83	1.10	55.19	58.90	7.71	1.01

Appendix

Table 10: RouteGate full metrics by resource tier (Belebele, LLaMA). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				TEN				RouteGate
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	34.65	38.78	11.02	1.77	31.29	39.07	13.92	1.78	35.92	34.28	5.42	1.91
Mid	25.64	33.70	13.13	1.91	30.75	31.89	10.09	1.96	27.95	36.62	3.38	2.07
Low	23.47	29.72	10.02	1.99	31.10	30.39	13.04	2.00	28.62	37.57	4.00	2.06

Appendix

Table 11: RouteGate full metrics by resource tier (MMLU-ProX-Lite, LLaMA). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				TEN				RouteGate
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	80.28	82.77	9.14	0.55	84.21	83.88	8.87	0.52	83.77	84.70	5.19	0.52
Mid	66.28	63.51	10.67	0.95	72.84	73.38	9.98	0.73	69.33	79.72	10.76	0.61
Low	45.95	45.67	9.42	1.22	56.44	59.08	10.92	0.99	71.14	62.78	13.19	0.96

Appendix

Table 12: RouteGate full metrics by resource tier (Belebele, Qwen). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				TEN				RouteGate
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	37.78	45.37	13.83	1.71	36.83	42.84	13.28	1.75	42.04	46.45	8.44	1.65
Mid	32.67	40.35	13.82	1.79	34.66	41.77	14.37	1.77	41.78	47.28	11.09	1.61
Low	28.57	38.15	12.81	1.83	34.86	41.48	16.39	1.78	38.04	47.87	10.65	1.61

Appendix

Table 13: RouteGate full metrics by resource tier (MMLU-ProX-Lite, Qwen). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				TEN				SeqGate
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	81.26	76.58	9.15	0.64	84.79	78.05	8.43	0.62	82.58	78.52	6.52	0.61
Mid	68.88	60.01	14.05	1.00	72.87	64.70	11.99	0.91	67.68	66.93	10.62	0.86
Low	44.59	39.38	12.36	1.29	50.68	47.61	6.47	1.18	69.70	66.13	11.34	0.88

Appendix

Table 14: SeqGate full metrics by resource tier (Belebele, LLaMA). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				TEN				SeqGate
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	34.65	38.78	11.02	1.77	32.16	35.29	12.51	1.88	33.76	34.83	10.59	1.88
Mid	25.64	33.70	13.13	1.91	24.67	27.17	13.17	2.06	33.73	24.61	8.76	2.12
Low	23.47	29.72	10.02	1.99	34.12	28.46	8.83	2.04	34.12	25.32	9.16	2.10

Appendix

Table 15: SeqGate full metrics by resource tier (MMLU-ProX-Lite, LLaMA). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				TEN				SeqGate
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	80.28	82.77	9.14	0.55	81.47	83.41	8.17	0.52	88.54	83.36	7.69	0.54
Mid	66.28	63.51	10.67	0.95	69.14	68.48	8.74	0.84	62.96	71.76	11.68	0.77
Low	45.95	45.67	9.42	1.22	49.32	51.00	10.70	1.13	83.64	71.71	8.35	0.78

Appendix

Table 16: SeqGate full metrics by resource tier (Belebele, Qwen). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Resource	T				TEN				SeqGate
	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$	Acc $\uparrow$	Conf $\uparrow$	ECE $\downarrow$	Ent $\downarrow$
High	37.78	45.37	13.83	1.71	41.00	44.76	17.77	1.71	47.13	42.47	10.41	1.77
Mid	32.67	40.35	13.82	1.79	38.70	44.01	13.52	1.71	51.64	37.57	14.04	1.87
Low	28.57	38.15	12.81	1.83	36.96	45.04	17.72	1.69	51.36	37.89	13.49	1.86

Appendix

Table 17: SeqGate full metrics by resource tier (MMLU-ProX-Lite, Qwen). Bold denotes the best value within each resource-tier row for each metric (higher Acc/Conf, lower ECE/Ent).

Panel of Figure 7 — Figure 7: Layer-wise accuracy under question translation.

Panel of Figure 8 — Figure 8: Layer-wise confidence under question translation.

Panel of Figure 9 — Figure 9: Layer-wise ECE under question translation.

Panel of Figure 10 — Figure 10: Layer-wise entropy under question translation.

Panel of Figure 11 — Figure 11: Layer-wise accuracy under choice translation.

Panel of Figure 12 — Figure 12: Layer-wise confidence under choice translation.

Panel of Figure 13 — Figure 13: Layer-wise ECE under choice translation.

Panel of Figure 14 — Figure 14: Layer-wise entropy under choice translation.

Figure 15: Root-cause composition by resource tier (H/M/L) across text scopes for MMLU-ProX-Lite, LLaMA. Bars show percent of outcomes per tier with labels; colors denote failure categories and stable success.

Figure 16: Root-cause composition by resource tier (H/M/L) across text scopes for MMLU-ProX-Lite, Qwen. Bars show percent of outcomes per tier with labels; colors denote failure categories and stable success.

Figure 17: Root-cause composition by resource tier (H/M/L) across text scopes for Belebele, Qwen. Bars show percent of outcomes per tier with labels; colors denote failure categories and stable success.

Model	Dataset	Tier	#Langs	$N$	RouteGate $\Delta$ Acc vs. T [95% CI]	SeqGate $\Delta$ Acc vs. T [95% CI]
LLaMA	Belebele	High	4	317	+2.32 [-3.60, +8.24]	+1.32 [-4.67, +7.31]
LLaMA	Belebele	Mid	3	239	+4.31 [-3.81, +12.43]	-1.20 [-9.54, +7.14]
LLaMA	Belebele	Low	2	158	+10.60 [-0.36, +21.56]	+25.11 [+14.55, +35.67]
LLaMA	MMLU-ProX-Lite	High	4	189	+1.27 [-8.36, +10.90]	-0.89 [-10.45, +8.67]
LLaMA	MMLU-ProX-Lite	Mid	3	146	+2.31 [-7.85, +12.47]	+8.09 [-2.35, +18.53]
LLaMA	MMLU-ProX-Lite	Low	2	92	+5.15 [-7.51, +17.81]	+10.65 [-2.34, +23.64]

Appendix

Table 18: Accuracy differences relative to T with 95% normal-approximation confidence intervals. Bold intervals exclude zero.

Variant	High-Res. err/AUROC	Mid-Res. err/AUROC	Low-Res. err/AUROC	err average	AUROC average
RF only	0.43 / 0.94	0.57 / 0.86	$-0.15$ / 0.32	0.28	0.71
ECE only	0.48 / 0.91	0.27 / 0.68	0.80 / 1.00	0.51	0.86
Entropy only	0.62 / 0.91	0.45 / 0.79	0.58 / 0.76	0.55	0.82
RF $\times$ ECE	0.49 / 0.96	0.40 / 0.79	0.42 / 0.68	0.44	0.81
RF $\times$ Entropy	0.62 / 1.00	0.64 / 0.89	$-0.07$ / 0.36	0.40	0.75
ECE $\times$ Entropy	0.56 / 0.91	0.31 / 0.71	0.56 / 0.84	0.48	0.82

Appendix

Table 19: Ablation study of the Risk Index for LLaMA on Belebele. Each resource-tier entry reports Spearman correlation with target-only error / AUROC for detecting groups with above-median target-only error.

Model / Dataset	Top 25%	Top 50%	Top 75%	Top 100%
LLaMA / Belebele	+3.27 / $-0.07$	+6.54 / $-0.34$	+10.73 / $-1.00$	+10.93 / $-0.71$
LLaMA / MMLU-ProX-Lite	+1.27 / +0.63	+1.45 / +0.11	+4.45 / +0.54	+4.41 / +0.27
Qwen / Belebele	+2.12 / +1.22	+3.62 / +0.86	+5.44 / +0.52	+5.43 / +0.87
Qwen / MMLU-ProX-Lite	+0.80 / +1.99	+0.92 / +2.27	+1.43 / +1.94	+10.54 / +1.49

Appendix

Table 20: SeqGate Risk Index threshold sensitivity. Each cell reports

\Delta

Acc /

\Delta

ECE versus T.

Model	Dataset	Tier	RI threshold	RouteGate routed	SeqGate routed	RouteGate cost ( $\times$ T)	SeqGate cost ( $\times$ T)
LLaMA	Belebele	High	0.500	19.8%	35.1%	1.35	1.75
LLaMA	Belebele	Mid	0.315	37.8%	55.0%	1.62	2.00
LLaMA	Belebele	Low	0.115	83.7%	90.3%	2.43	2.00
LLaMA	MMLU-ProX-Lite	High	0.500	3.3%	40.0%	1.03	2.00
LLaMA	MMLU-ProX-Lite	Mid	0.315	15.5%	60.0%	1.04	2.00
LLaMA	MMLU-ProX-Lite	Low	0.115	35.8%	89.0%	1.06	2.00

Appendix

Table 21: Tier-level routing rates and normalized LLM inference cost relative to T. Translation-system cost and latency are not included.

Cross-Lingual Reasoning in Language Models

Cross-lingual reasoning in language models studies how models solve problems and transfer knowledge across languages, scripts, and cultures.

See all 69 papers

Learning When to Translate for Multilingual Reasoning
Jun 1, 2026Deokhyung Kang, Hyounghun KimMultilingual Language Models Cross-Lingual Reasoning in Language Models
Graduate School of Artificial Intelligence · Department of Computer Science & Engineering
The Reliance Hypothesis: Chain-of-Thought Faithfulness Rises as Language Resource Level Falls
Oct 16, 2026Lucas Resck, Jingyi Sun, Qianli Wang +2CoT Faithfulness Faithfulness of Language Model Reasoning
Language Technology Lab, University of Cambridge · University of Copenhagen · Technische Universität Berlin +1
ReasonXL: A Large-Scale Corpus for Language-Specific Reasoning in LLMs
Apr 14, 2026Daniil Gurgurov, Tom Röhr, Sebastian von Rohrscheidt +3Cross-Lingual Reasoning in Language Models Multilingual Language Models
German Research Center for Artificial Intelligence (DFKI), Saarland University · Berliner Hochschule für Technik

Jun 1, 2026J/K move · Enter open · S save

Learning When to Translate for Multilingual Reasoning

Deokhyung Kang, Hyounghun Kim

Graduate School of Artificial Intelligence · Department of Computer Science & Engineering

Multilingual Language Models Cross-Lingual Reasoning in Language Models RL for Language Model Reasoning Low-Resource Language Processing

Open paper AskPDF arXiv

Reasoning language models (RLMs) achieve strong performance on complex reasoning tasks, but still exhibit substantial multilingual reasoning gaps, largely due to language-understanding failures when interpreting non-English inputs. Translation into English can mitigate these failures by expressing non-English inputs in a form that RLMs can more reliably interpret, yet translating every input incurs unnecessary inference cost when the model can reason reliably from the original query. To address this challenge, we propose Luar, a Language Understanding Boundary-aware Reinforcement Learning framework that trains RLMs to selectively invoke translation when their direct understanding is unreliable. Luar trains the model to either solve the original input directly or reason over its English translation, encouraging translation only when translator-augmented reasoning is expected to substantially outperform direct reasoning. Across multilingual reasoning benchmarks, Luar outperforms training-based baselines, with particularly large gains on low-resource languages. On Qwen3-4B, it improves average accuracy by 8.4 percentage points over standard GRPO. Further analysis shows that Luar concentrates translator calls on queries where translation benefits reasoning, helping to reduce inference cost. The project will be made publicly available at https://github.com/deokhk/LUAR.

Abstract

Cross-Lingual Reasoning in Language Models

Explore similar work

Learning When to Translate for Multilingual Reasoning

The Reliance Hypothesis: Chain-of-Thought Faithfulness Rises as Language Resource Level Falls

ReasonXL: A Large-Scale Corpus for Language-Specific Reasoning in LLMs