Recovered in Translation: Efficient Pipeline for Automated Translation of Benchmarks and Datasets

Hanna Yukhymenko; Anton Alexandrov; Martin Vechev

cs.CLFeb 25, 2026

Recovered in Translation: Efficient Pipeline for Automated Translation of Benchmarks and Datasets

Authors: Hanna Yukhymenko, Anton Alexandrov, Martin Vechev

Organizations: INSAIT, Sofia University "St. Kliment Ohridski" · ETH Zurich

Abstract

The reliability of multilingual Large Language Model (LLM) evaluation is currently compromised by the inconsistent quality of translated benchmarks. Existing resources often suffer from semantic drift and context loss, which can lead to misleading performance metrics. In this work, we present a fully automated framework designed to address these challenges by enabling scalable, high-quality translation of datasets and benchmarks. We demonstrate that adapting test-time compute scaling strategies, specifically Universal Self-Improvement (USI) and our proposed multi-round ranking method, T-RANK, allows for significantly higher quality outputs compared to traditional pipelines. Our framework ensures that benchmarks preserve their original task structure and linguistic nuances during localization. We apply this approach to translate popular benchmarks and datasets into eight Eastern and Southern European languages (Ukrainian, Bulgarian, Slovak, Romanian, Lithuanian, Estonian, Turkish, Greek). Evaluations using both reference-based metrics and LLM-as-a-judge show that our translations surpass existing resources, resulting in more accurate downstream model assessment. We release both the framework and the improved benchmarks to facilitate robust and reproducible multilingual AI development.

Figures & tables

Figure 1: Universal Self-Improvement (USI) method workflow outlook

Figure 2: Translation Ranking (T-RANK) method workflow outlook

Method	WMT24++	FLORES
Baseline	0.827	0.937
SC (with check)	0.821	0.937
Best-of-N (n=5)	0.843	0.943
USI (n=5)	0.843	0.945
T-RANK (p=5)	0.845	0.940

Table 1: COMET (reference-based) translation quality scores for WMT24++ and FLORES EN

\to

UK pair (GPT-4o-mini).

n

denotes number of sampled candidates,

p

denotes number of different translation prompts (for

p>1

we use

n=1

Language	Method	Wins	Draws	Losses
UK	T-RANK	8750	3276	2016
RO	T-RANK	8376	3020	2646
LT	USI	7382	3181	3478

Table 2: LLM-as-a-judge comparison between Global-MMLU and our translations (T-RANK/USI) using Gemini-2.5-Flash as judge. Our translations win significantly more comparisons across all evaluated languages.

Benchmark	Languages	AVG $\Delta$
Winogrande	UK, RO, SK, BG, LT, TR, EL, ET	+3.42%
ARC-Challenge	UK, RO, SK, BG	+2.35%
Hellaswag	UK, RO, SK, BG	+1.63%
MMLU	UK, RO, SK, BG, LT, TR, EL	+0.94%

Table 3: Average improvement per benchmark across all languages

Language	AVG $\Delta$	Language	AVG $\Delta$
Greek	+3.89%	Romanian	+2.09%
Ukrainian	+2.7%	Slovak	+1.66%
Turkish	+2.65%	Estonian	+1.63%
Lithuanian	+2.6%	Bulgarian	+1.37%

Table 4: Average improvement per language across all benchmarks

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 3 — Figure 3: Translation Ranking (T-RANK) Correction Example

Panel of Figure 4 — Figure 4: Universal Self-Improvement (USI) Correction Example

Input position	Avg. rank	Top-5 rank combination (positions 1–5)	Occurrences
1	3.01	(4, 1, 3, 2, 5)	4731
2	2.06	(3, 1, 2, 4, 5)	3141
3	2.93	(4, 1, 2, 3, 5)	2997
4	3.07	(3, 1, 4, 2, 5)	2600
5	3.93	(2, 1, 3, 4, 5)	2428
Equal ranks		6324

Appendix

Table 5: Positional bias in T-RANK ranking for MMLU (EN

\to

UK,

n{=}5

candidates for

p{=}1

, translation model gpt-4o-mini ). Average rank per input position (lower = ranked better) and the five most frequent rank combinations. Position 2 receives the best average rank (2.06) and holds rank 1 in every top combination, confirming a residual bias towards the second candidate despite the multi-round rotation strategy.

✗ Bad Example (MuBench, Answer Leakage)	✓ Good Example (Ours, Correct Translation)
Вони хвилювалися, що вино зпсу лжко та ковдру, але _ не було зпсовано .	Вони хвилювалися, що вино зпсу лжко та ковдру, але _ не бу(-в/-ла/-ло/-ли) зпсовано.
What does the blank _ refer to?	What does the blank _ refer to?
Option A: ковдра	Option A: ковдра
Option B: лжко	Option B: лжко
Answer with A or B.	Answer with A or B.
Answer: лжко	Answer: лжко ?

Appendix

Table 6: Winogrande translation examples. Left (MuBench UK Winogrande): answer leakage where the correct answer “ лжко ” is implied grammatically in the question itself. Right (our translation): correct translation with verb morphology masking “ бу(-в/-ла/-ло/-ли) ” to prevent answer leakage.

Issue Type	Example	Impact
Semantic Drift	Original: “relatively long lifespan” Translated: ‘‘життвий цикл’’ (life cycle) Should be: ‘‘тривалсть життя’’ (lifespan)	Changes the question meaning; “life cycle” refers to developmental stages, not longevity
Wrong Terminology	Original: “aquatic organism” Translated: ‘‘водяний органзм’’ Should be: ‘‘водний органзм’’	‘‘Водяний’’ means “watery” (like soup), while ‘‘водний’’ is the correct scientific term for aquatic
Grammar Errors	Original: “parental care for offspring” Translated: ‘‘турботи за потомством’’ Should be: ‘‘турботи про потомство’’	Incorrect preposition usage creates unnatural phrasing that may confuse native speakers
Literal Translation	Original: “Only I” Translated: ‘‘Тльки я’’ Should be: ‘‘Тльки I’’	The Latin numeral “I” is mis-read as the letter and rendered as the Ukrainian pronoun ‘‘я’’ (I/me); the Roman numeral must be preserved in Latin script

Appendix

Table 7: MMLU translation errors (Ukrainian) found in existing benchmarks (Global-MMLU, Okapi). Four recurring error categories with concrete examples and their impact on evaluation reliability.

✗ Bad Example (Okapi, Literal Translation)	✓ Good Example (Ours, Natural Phrasing)
Ми бачимо початковий екран заголовка. Ми бачимо двчинку бжать та робить високий стрибок, пройшовши через перекладину . Ми	Ми бачимо титульний екран. Ми бачимо, як двчина бжить , викону стрибок у висоту перелазить через планку . Ми
Яке закнчення ма найбльший сенс?	Яке закнчення ма найбльший сенс?
A) бачимо близько 30 двчат стрибають.	A) бачимо, як близько 30 двчат стрибають.
B) бачимо, як вона закнчу перекладину та викону рутину.	B) бачимо, як вона закнчу стрибок над планкою забира свою рутину.
C) потм ми бачимо повтор повльний повтор.	C) потм бачимо повтор уповльнений повтор.
D) бачимо, як двчина кида блий м’яч, вн летить прямо по полю.	D) бачимо, як двчина кида блий м’яч, вн летить прямо через поле.

Appendix

Table 8: Hellaswag translation example (Ukrainian). Original English: “We see an opening title screen. We see a girl run and perform a high jump and make it over the bar. We… — which ending makes the most sense?”

Benchmark	Train	Val	Dev	Test	All
MMLU	–	1531	285	14042	15858
Winogrande	–	1267	–	–	–
ARC-Challenge	1119	299	–	1172	2291
Hellaswag	–	–	–	10042	10042
Total	1119	3097	285	25556	29757

Appendix

Table 9: Number of examples in each benchmark split (– indicates split not available or not used).

Benchmark	UK	SK	RO	LT	ET	BG	EL	TR
MMLU	4o/TR	4o/TR	4o/TR	4o/USI	4o/USI	4o/TR	4o/USI	4o/USI
Hellaswag	4o	4o	4o	G2F	G2F	G2F	G2F	G2F
ARC	4o	G2F	G2F	G2F	G2F	G2F	G2F	G2F
Winogrande	4o	G2F	G2F	G2F	G2F	G2F	G2F	G2F

Appendix

Table 10: Model and method for benchmark translation. 4o = GPT-4o-mini, G2F = Gemini-2.0-Flash, TR = T-RANK. If method not specified, USI was used.

	Hellaswag			Winogrande			ARC-Challenge			MMLU
Model	O	Ot	$\Delta$ %	O	Ot	$\Delta$ %	O	Ot	$\Delta$ %	O	Ot	$\Delta$ %
Gemma-3-12B-IT	0.687	0.625	+6.2%	0.580	0.619	-3.9%	0.517	0.470	+4.8%	0.614	0.603	+1.2%
Llama-3.1-8B	0.570	0.540	+3%	0.549	0.495	+5.5%	0.431	0.416	+1.5%	0.497	0.489	+0.9%
Gemma-3-4B-IT	0.578	0.528	+5%	0.540	0.501	+3.9%	0.453	0.417	+3.7%	0.453	0.444	+1%
Qwen3-8B-IT	0.540	0.520	+2%	0.569	0.546	+2.3%	0.448	0.404	+4.4%	0.619	0.599	+2%

Appendix

Table 11: Ukrainian. O=Ours, Ot=Other (Okapi/MuBench/Global-MMLU)

	Hellaswag			Winogrande			ARC-Challenge			MMLU
Model	O	Ot	$\Delta$ %	O	Ot	$\Delta$ %	O	Ot	$\Delta$ %	O	Ot	$\Delta$ %
Gemma-3-12B-IT	0.678	0.681	-0.2%	0.661	0.589	+7.2%	0.486	0.454	+3.2%	0.628	0.614	+1.4%
Gemma-3-4B-IT	0.567	0.567	0%	0.575	0.531	+4.5%	0.412	0.394	+1.8%	0.479	0.473	+0.6%
Llama-3.1-8B-IT	0.576	0.581	-0.5%	0.616	0.521	+9.5%	0.358	0.357	+0.2%	0.529	0.519	+0.9%
Qwen3-8B-IT	0.538	0.537	+0.1%	0.588	0.566	+2.3%	0.368	0.359	+0.8%	0.650	0.632	+1.8%

Appendix

Table 12: Romanian. O=Ours, Ot=Other (Okapi/MuBench/Global-MMLU)

	Winogrande			ARC-Challenge
Model	Ours	Other	$\Delta$ %	Ours	Other	$\Delta$ %
Gemma-3-12B-IT	0.610	0.589	+2.1%	0.539	0.509	+3%
Gemma-3-4B-IT	0.554	0.517	+3.7%	0.448	0.444	+0.4%
Llama-3.1-8B-IT	0.547	0.515	+3.2%	0.443	0.403	+4%
Qwen3-8B-IT	0.555	0.559	-0.4%	0.428	0.427	+0.1%

Appendix

Table 13: Slovak. Other=MuBench/Okapi

	Winogrande			MMLU
Model	Ours	Other	$\Delta$ %	Ours	Other	$\Delta$ %
Gemma-3-12B-IT	0.631	0.557	+7.3%	0.585	0.573	+1.3%
Gemma-3-4B-IT	0.530	0.506	+2.3%	0.433	0.410	+2.3%
Llama-3.1-8B-IT	0.527	0.498	+2.9%	0.431	0.417	+1.4%
Qwen3-8B-IT	0.551	0.530	+2.1%	0.553	0.541	+1.2%

Appendix

Table 14: Lithuanian. Other=MuBench/Global-MMLU

	Winogrande
Model	Ours	Other	$\Delta$ %
Gemma-3-12B-IT	0.615	0.566	+4.9%
Gemma-3-4B-IT	0.541	0.516	+2.5%
Llama-3.1-8B-IT	0.543	0.509	+3.4%
Qwen3-8B-IT	0.517	0.503	+1.4%

Appendix

Table 15: Estonian. Other=MuBench

	Winogrande			MMLU
Model	Ours	Other	$\Delta$ %	Ours	Other	$\Delta$ %
Gemma-3-12B-IT	0.635	0.571	+6.5%	0.587	0.581	+0.7%
Gemma-3-4B-IT	0.583	0.514	+7%	0.447	0.437	+1%
Llama-3.1-8B-IT	0.573	0.518	+5.5%	0.494	0.480	+1.4%
Qwen3-8B-IT	0.545	0.571	-2.7%	0.588	0.569	+1.8%

Appendix

Table 16: Turkish. Other=MuBench/Global-MMLU

	Winogrande			MMLU
Model	Ours	Other	$\Delta$ %	Ours	Other	$\Delta$ %
Gemma-3-12B-IT	0.657	0.601	+5.6%	0.593	0.580	+1.3%
Gemma-3-4B-IT	0.598	0.518	+8.0%	0.428	0.421	+0.7%
Llama-3.1-8B-IT	0.593	0.520	+7.3%	0.465	0.446	+2%
Qwen3-8B-IT	0.581	0.528	+5.3%	0.563	0.553	+1.1%

Appendix

Table 17: Greek. Other=MuBench/Global-MMLU

	Hellaswag			Winogrande			ARC-Challenge			MMLU
Model	Ours	Other	$\Delta$ %	Ours	Other	$\Delta$ %	Ours	Other	$\Delta$ %	Ours	Other	$\Delta$ %
Gemma-3-12B-IT	0.708	0.690	+1.8%	0.643	0.677	-3.5%	0.526	0.495	+3.1%	0.605	0.619	-1.3%
Gemma-3-4B-IT	0.590	0.569	+2.2%	0.588	0.595	-0.7%	0.430	0.410	+2%	0.454	0.469	-1.5%
Llama-3.1-8B-IT	0.539	0.528	+1.1%	0.575	0.605	-3.1%	0.394	0.367	+2.7%	0.481	0.487	-0.6%
Qwen3-8B-IT	0.536	0.518	+1.8%	0.602	0.629	-2.7%	0.414	0.394	+2%	0.619	0.617	+0.2%

Appendix

Table 18: Bulgarian. Other=INSAIT (professional translators)

	Winogrande
Model	Ours	Other	$\Delta$ %
Gemma-3-12B-IT	0.643	0.597	+4.6%
Gemma-3-4B-IT	0.588	0.506	+8.2%
Llama-3.1-8B-IT	0.575	0.505	+6.9%
Qwen3-8B-IT	0.602	0.559	+4.3%

Appendix

Table 19: Bulgarian. Other=MuBench

Method	EN $\to$ UK	EN $\to$ SK	EN $\to$ RO	EN $\to$ LT	EN $\to$ ET	EN $\to$ BG	EN $\to$ TR	EN $\to$ EL
Baseline	0.726	0.741	0.882	0.741	0.788	0.751	0.718	0.802
SC	0.708	0.725	0.869	0.735	0.788	0.749	0.715	0.793
BoN n=5	0.743	0.756	0.895	0.767	0.809	0.788	0.736	0.819
USI n=5	0.750	0.759	0.899	0.766	0.806	0.791	0.733	0.817
T-RANK n=5	0.739	0.745	0.885	0.749	0.799	0.776	0.729	0.811
USI multi-prompt p=5	0.755	0.764	0.898	0.771	0.809	0.793	0.736	0.825

Appendix

Table 20: WMT QE (Quality Estimation) reference-free COMET scores for GPT-4o-mini

Method	EN $\to$ UK	EN $\to$ SK	EN $\to$ RO	EN $\to$ LT	EN $\to$ ET	EN $\to$ BG	EN $\to$ TR	EN $\to$ EL
Baseline	0.827	0.822	0.873	0.788	0.821	0.834	0.776	0.820
SC	0.821	0.817	0.869	0.790	0.822	0.834	0.834	0.821
BoN n=5	0.844	0.837	0.884	0.805	0.838	0.852	0.791	0.836
USI n=5	0.843	0.839	0.888	0.807	0.842	0.856	0.791	0.838
T-RANK n=5	0.840	0.834	0.885	0.803	0.836	0.855	0.793	0.837
USI multi-prompt p=5	0.849	0.847	0.891	0.817	0.848	0.862	0.803	0.844

Appendix

Table 21: WMT COMET reference-based scores for GPT-4o-mini

Method	EN $\to$ UK	EN $\to$ SK	EN $\to$ RO	EN $\to$ LT	EN $\to$ ET	EN $\to$ BG	EN $\to$ TR	EN $\to$ EL
Baseline	0.904	0.906	0.947	0.906	0.927	0.903	0.904	0.896
SC	0.902	0.908	0.950	0.911	0.935	0.902	0.903	0.896
BoN n=5	0.910	0.917	0.956	0.921	0.947	0.916	0.912	0.911
USI n=5	0.913	0.915	0.956	0.924	0.943	0.917	0.910	0.909
T-RANK n=5	0.899	0.906	0.952	0.909	0.932	0.911	0.903	0.902
USI multi-prompt p=5	0.919	0.921	0.961	0.932	0.948	0.922	0.918	0.912

Appendix

Table 22: FLORES QE (Quality Estimation) reference-free COMET scores for GPT-4o-mini

Method	EN $\to$ UK	EN $\to$ SK	EN $\to$ RO	EN $\to$ LT	EN $\to$ ET	EN $\to$ BG	EN $\to$ TR	EN $\to$ EL
Baseline	0.937	0.938	0.939	0.906	0.894	0.943	0.920	0.920
SC	0.937	0.937	0.936	0.911	0.902	0.940	0.916	0.923
BoN n=5	0.943	0.945	0.945	0.917	0.918	0.951	0.927	0.931
USI n=5	0.945	0.942	0.947	0.922	0.919	0.951	0.924	0.932
T-RANK n=5	0.938	0.937	0.943	0.910	0.903	0.947	0.920	0.929
USI multi-prompt p=5	0.947	0.946	0.949	0.928	0.918	0.952	0.931	0.934

Appendix

Table 23: FLORES COMET reference-based scores for GPT-4o-mini

Method	EN $\to$ UK	EN $\to$ SK	EN $\to$ RO	EN $\to$ LT	EN $\to$ ET	EN $\to$ BG	EN $\to$ TR	EN $\to$ EL
Baseline	0.688	0.703	0.777	0.704	0.737	0.708	0.683	0.719
SC	0.687	0.706	0.777	0.706	0.736	0.709	0.685	0.722
BoN n=5	0.687	0.707	0.780	0.710	0.743	0.713	0.689	0.723
USI multi-prompt p=5	0.698	0.716	0.785	0.718	0.753	0.721	0.697	0.731
T-RANK multi-prompt p=5	0.697	0.713	0.781	0.716	0.749	0.718	0.695	0.726

Appendix

Table 24: WMT QE (Quality Estimation) reference-free COMET scores for Gemini-2.0-flash

Method	EN $\to$ UK	EN $\to$ SK	EN $\to$ RO	EN $\to$ LT	EN $\to$ ET	EN $\to$ BG	EN $\to$ TR	EN $\to$ EL
Baseline	0.828	0.829	0.867	0.805	0.847	0.840	0.778	0.821
SC	0.826	0.834	0.868	0.808	0.845	0.841	0.779	0.824
BoN n=5	0.828	0.836	0.870	0.814	0.852	0.843	0.785	0.823
USI multi-prompt p=5	0.841	0.843	0.881	0.826	0.862	0.856	0.797	0.836
T-RANK multi-prompt p=5	0.841	0.849	0.882	0.827	0.861	0.859	0.800	0.838

Appendix

Table 25: WMT COMET reference-based scores for Gemini-2.0-flash

Method	Prompt Template
Base Translation	Instructions: Imagine you’re part of a team at an international education center that’s revamping its exams for a global audience. Your job is to translate an English question and its answer options into <target_language> so that students from <target_language> schools can be evaluated too. Just provide the final translation—leave out any extra comments or explanations. Use language which is authentic for <target_language> natives. Remember to keep the answer options connected to the question, using the same format as the original (a list for multiple choices or plain text for a single answer). Please do not translate valid code in any of the programming languages. Original text: {"Original_question": "<original_question>", "Original_answers": "<original_answers>"} Output instructions: Now, please give your final translation in <target_language> exactly in this format, with only the translated content: {"Question": "your_translated_question", "Answers": "translated_answers"}
USI Judge	My task is to translate BENCHMARK questions with answers from English to <target_language>. Your task is to evaluate if the response preserves the original question idea and to verify the correctness of declension and conjunction of words in the target language. The original text in English is: Question: <original_question>, Answers: <original_answers> I have generated the following responses: <responses> Combine the best features from responses to form the best response from grammatical and coherent points of view. Look for the next metrics: * Quality of translation, including grammatical correctness * Domain knowledge - were the terms correctly translated w.r.t to domain? Were coding terms or function names preserved (not translated)? * Is the question text fully translated? Is the question idea preserved? * Are all the answer options/texts fully translated? * Are the words written correctly? Are there any typos? Output only the selected response: Question: selected question, Answer: selected answers
T-RANK Ranking	My task is to translate BENCHMARK questions from English to <target_language>. Your task is to rank my translations and select the best one. Ranking criteria: * Quality of translation * Domain knowledge - were the terms correctly translated w.r.t to domain? Were coding terms or function names preserved (one would usually not translate coding operators or function names)? * Is the question text fully translated? Is the question idea preserved? * Are all the answer options/texts fully translated? * Are the words written correctly? Are there any typos? * Are the declension and the conjunction of words written correctly in the target language if you connect answer options with the question? Original: {"original_question": "<original_question>", "original_answers": <original_answers>} Candidates: <responses> Instruction: Select the best response (1st place = best). Correct if needed before output. Output: Reasoning, then: {"summary": "…", "final_ranks": {…}, "rankings_list": […], "best_translation": {…}}
Best-of-N Scoring	My task is to translate questions from English to <target_language>. Score my translations 1-10 (10 = best). Original : Question: <original_question>, Answers: <original_answers> Scoring metrics : (1) Translation quality; (2) Question fully translated?; (3) All answers translated?; (4) Question idea preserved?; (5) Correct grammar? Responses : <responses> Output scores only: Response 1: score, Response 2: score, …, Answers: [list of scores]

Appendix

Table 26: Translation prompt templates for different methods

Panel of Figure 5 — Figure 5: LLM-as-a-judge Quality Evaluation Prompt

Explore similar work

Diagnosing Translated Benchmarks: An Automated Quality Assurance Study of the EU20 Benchmark Suite
Apr 2, 2026Klaudia Thellmann, Bernhard Stadler, Michael FärberBenchmark Auditing MT Evaluation
TU Dresden and ScaDS.AI · InfAI e.V.
Quantifying the Impact of Translation Errors on Multilingual LLM Evaluation
May 24, 2026Klaudia-Doris Thellmann, Bernhard Stadler, Michael Färber +1Multilingual Language Model Evaluation MT Evaluation
TUD Dresden University of Technology · 2InfAI e.V. · 3Amazon +1
Scaling Multilingual Safety and Reasoning Evaluation through Reasoning-Guided Translation
Oct 14, 2026Zahraa Al Sahili, Ioannis Patras, Matthew PurverMultilingual Language Model Evaluation Language Model Safety Evaluation
School of Electronic Engineering and Computer Science, Queen Mary University of London, London, UK · Department of Knowledge Technologies, Jožef Stefan Institute, Ljubljana, Slovenia

Apr 2, 2026J/K move · Enter open · S save

Diagnosing Translated Benchmarks: An Automated Quality Assurance Study of the EU20 Benchmark Suite

Klaudia Thellmann, Bernhard Stadler, Michael Färber

TU Dresden and ScaDS.AI · InfAI e.V.

Benchmark Auditing MT Evaluation Automated Evaluation Multilingual Language Model Evaluation

Open paper AskPDF arXiv

Machine-translated benchmark datasets reduce costs and offer scale, but noise, loss of structure, and uneven quality weaken confidence. What matters is not merely whether we can translate, but also whether we can measure and verify translation reliability at scale. We study translation quality in the EU20 benchmark suite, which comprises five established benchmarks translated into 20 languages, via a three-step automated quality assurance approach: (i) a structural corpus audit with targeted fixes; (ii) quality profiling using a neural metric (COMET, reference-free and reference-based) with translation service comparisons (DeepL / ChatGPT / Google); and (iii) an LLM-based span-level translation error landscape. Trends are consistent: datasets with lower COMET scores exhibit a higher share of accuracy/mistranslation errors at span level (notably HellaSwag; ARC is comparatively clean). Reference-based COMET on MMLU against human-edited samples points in the same direction. We release cleaned/corrected versions of the EU20 datasets, and code for reproducibility. In sum, automated quality assurance offers practical, scalable indicators that help prioritize review -- complementing, not replacing, human gold standards.

Abstract

Explore similar work

Diagnosing Translated Benchmarks: An Automated Quality Assurance Study of the EU20 Benchmark Suite

Quantifying the Impact of Translation Errors on Multilingual LLM Evaluation

Scaling Multilingual Safety and Reasoning Evaluation through Reasoning-Guided Translation