Paper deep dive
Discovering Translation-Worthy Languages with E-Values
Wajdi Ben Saad, Safa Madiouni
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 9/12/2026, 2:47:37 AM
Summary
This paper introduces a language-level router for multilingual text classification that uses paired e-processes to statistically determine when translation improves classification accuracy. By monitoring discordant pairs between direct classification and translation-assisted classification, the method identifies specific languages or locales where translation is beneficial, freezing the routing policy after sufficient evidence is accumulated. Evaluated on SIB-200 and MASSIVE datasets, the router significantly improves held-out accuracy over direct classification while maintaining statistical validity and audibility.
Entities (11)
Relation Signals (10)
Paired E-Process Router â evaluatedon â SIB-200
confidence 98% · On SIB-200 and MASSIVE, the router selects translation...
Paired E-Process Router â evaluatedon â MASSIVE
confidence 98% · On SIB-200 and MASSIVE, the router selects translation...
Paired E-Process Router â selectstranslationfor â Amharic
confidence 95% · SIB-200 selects Swahili, Bengali, Telugu, and Amharic.
Paired E-Process Router â selectstranslationfor â Swahili
confidence 95% · SIB-200 selects Swahili, Bengali, Telugu, and Amharic.
Paired E-Process Router â selectstranslationfor â Bengali
confidence 95% · SIB-200 selects Swahili, Bengali, Telugu, and Amharic.
Paired E-Process Router â selectstranslationfor â Telugu
confidence 95% · SIB-200 selects Swahili, Bengali, Telugu, and Amharic.
Paired E-Process Router â uses â E-values
confidence 95% · We introduce a language-level router based on paired e-processes... E-values are nonnegative evidence measures...
Paired E-Process Router â improvesaccuracyon â SIB-200
confidence 90% · improving held-out accuracy over direct classification by 8.14... percentage points
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Choosing when to translate multilingual documents is a central routing problem in text classification: translation can improve predictions for some languages while degrading others or adding unnecessary computation. Uniform translation and heuristic language tiers do not provide statistically controlled route selection. We introduce a language-level router based on paired e-processes that continuously compares direct and translation-assisted classification before freezing a routing policy. A familywise-controlled threshold of 280 bounds the probability of any false route across 14 eligible languages per dataset by 0.05. On SIB-200 and MASSIVE, the router selects translation for 4 of 15 languages and 14 of 15 locales, improving held-out accuracy over direct classification by 8.14 and 16.70 percentage points, respectively. All 28 decisions remain stable across 50 outcome-independent orderings and relative to the per-group threshold. Our results demonstrate that paired e-processes enable statistically controlled, anytime-valid, and auditable multilingual classification routing.
Tags
Links
- Source: https://arxiv.org/abs/2609.06593v1
- Canonical: https://arxiv.org/abs/2609.06593v1
Trouble viewing inline? Open PDF directly â
Full Text
26,620 characters extracted from source content.
Expand or collapse full text
Discovering Translation-Worthy Languages with E-Values Wajdi Ben Saad Carthago Labs Paris, France wajdi@carthagolabs.org Safa Madiouni IĂSEG Paris, France s.madiouni@ieseg.fr Abstract Choosing when to translate multilingual documents is a central routing problem in text classification: translation can improve predictions for some languages while degrading others or adding unnecessary computation. Uniform translation and heuristic language tiers do not provide statistically controlled route selection. We introduce a language-level router based on paired e-processes that continuously compares direct and translation-assisted classification before freezing a routing policy. A familywise-controlled threshold of 280 bounds the probability of any false route across 14 eligible languages per dataset by 0.05. On SIB-200 and MASSIVE, the router selects translation for 4 of 15 languages and 14 of 15 locales, improving held-out accuracy over direct classification by 8.14 and 16.70 percentage points, respectively. All 28 decisions remain stable across 50 outcome-independent orderings and relative to the per-group threshold. Our results demonstrate that paired e-processes enable statistically controlled, anytime-valid, and auditable multilingual classification routing. 1 Introduction and related work Inference systems increasingly choose between a default computation and a costlier intervention. When evidence is monitored as it accumulates, fixed-horizon tests do not justify data-dependent stopping. E-values are nonnegative evidence measures with null expectation at most one, and e- processes extend this principle to continuous monitoring under explicit sequential assumptions (Vovk and Wang, 2021; GrĂŒnwald et al., 2024; Ramdas et al., 2023). We study a new decision role for this established methodology: constructing a persistent group-level router from paired outcomes. Multilingual classification provides a concrete application. Inputs may be classified directly by a multilingual model or translated into English before classification. Translate-test performance varies with language, task, and pipeline (Artetxe et al., 2023), yet existing language-level routing uses hand-specified resource tiers (Ben Saad and Madiouni, 2026). Such tiers neither measure evidence that translation improves a group nor necessarily transfer across tasks. Recent multilingual and translation-routing systems learn instance-level strategy or model selection (Wu et al., 2026; Luo et al., 2026), while FrugalML and RouteLLM learn query-specific qualityâcost policies (Chen et al., 2020; Ong et al., 2025); our aim is different: assign a predefined group to the intervention only after paired held-out observations provide sufficient statistical evidence. Our construction follows the discordance principle behind McNemarâs paired test (McNemar, 1947). Among examples on which the two paths disagree in correctness, translation has higher marginal accuracy exactly when fixes are more likely than regressions. We monitor this Bernoulli sequence with an established mixture e-process for paired model evaluation (Choe and Ramdas, 2023; Kotawala, 2026). PACE uses paired anytime-valid evidence for sequential agent-update acceptance; here, crossings define a persistent group-level inference route (Shawn, 2026). All routes are frozen before test evaluation. arXiv:2609.06593v1 [cs.CL] 6 Sep 2026 We evaluate independently discovered routers on SIB-200 topic classification (Adelani et al., 2024) and MASSIVE intent classification (FitzGerald et al., 2023). Our contribution is not a new e-process, but (i) a paired-evidence formulation of group routing, (i) a discoveryâfreezeâtest protocol that turns crossings into an auditable policy, and (i) evidence that the resulting boundary is task-specific rather than a universal language tier. 2 Method 2.1 Sign-sufficient paired reduction For groupg, letDdenote a default path andTan intervention. On held-out discovery examplei, define correctness indicators C D gi = 1by D gi = y gi , C T gi = 1by T gi = y gi .(1) A discordant pair is a fix,(C D ,C T ) = (0, 1), or a regression,(1, 0). Write their probabilities asa g and b g , let d g = a g + b g be the discordance rate, and let q g = a g /d g . Then â acc,g = Acc T,g â Acc D,g = a g â b g = d g (2q g â 1).(2) Ford g > 0,q g > 1/2if and only ifThas higher marginal accuracy. Concordant pairs contribute equally to both accuracies, so discarding them loses no information about the sign of the paired accuracy difference. Ifd g = 0, the difference is zero and the group retainsD. The target is paired accuracy, not Macro-F1, whose nonlinear dependence on the multiclass confusion matrix requires a different process. 2.2 Discordance e-process In a prespecified order, letX g,t = 1for thetth fix and0for thetth regression, andS g,t = P t j=1 X g,j . We test H 0,g : q g †1/2 with the normalized uniform mixture E g,t = 2 Z 1 1/2 p S g,t (1â p) tâS g,t (1/2) t dp, E g,0 = 1.(3) The factor two normalizes a uniform mixing distribution on[1/2, 1]; this mixture is fixed before observing outcomes and requires no effect-size tuning. LetF g,tâ1 contain the prespecified order and the firsttâ 1discordant outcomes. Under i.i.d. paired sampling, the ordered discordant outcomes are Bernoulli with parameterq g . More generally, we assumeP (X g,t = 1 | F g,tâ1 ) †1/2under the null. For each fixedp > 1/2, this condition makes the corresponding likelihood-ratio process a nonnegative supermartingale; mixing overppreserves the e-process property (Vovk and Wang, 2021; Ramdas et al., 2023). 2.3 Anytime validity and evidence allocation For a prespecified testing levelα g , define the first crossingÏ g = inft : E g,t â„ 1/α g . The e-process maximal inequality gives Pr H 0,g (Ï g <â) = Pr H 0,g sup t E g,t â„ 1/α g †α g .(4) Thus the same threshold is valid under continuous inspection and stopping at the first crossing. Because the router makes simultaneous decisions for 14 translation-eligible groups within each dataset, we use the familywise error allocationα g = .05/14. The resulting e-value threshold is 1/α g = 280, which guarantees by the union bound that the probability of at least one false route is at most .05, without requiring independence across groups. We retain the per-group allocationα g = .05 (threshold 20) as a sensitivity analysis. 2.4 Evidence time, routing, and freezing The process advances in discordance time, whereas data arrive in example time. If a route crosses aftertdiscordances, approximatelyt/d g paired examples are required; hence evidence requirements 2 050100150200250300350400 Discordant pairs observed 0 25 50 75 100 125 150 175 Log e-process value SIB-200: complete discovery trajectories Afrikaans Amharic Arabic Bengali German English French Japanese Dutch Polish Spanish Swahili Telugu Turkish Chinese Primary FWER threshold: 280 Per-group sensitivity threshold: 20 0200400600800 Discordant pairs observed 0 100 200 300 400 500 Log e-process value MASSIVE: complete discovery trajectories Afrikaans Amharic Arabic Bengali German English Spanish French Dutch Polish Portuguese Romanian Swahili Telugu Turkish Primary FWER threshold: 280 Per-group sensitivity threshold: 20 Figure 1: Discovery e-processes for SIB-200 (left) and MASSIVE (right). The primary familywise- controlled threshold is 280; threshold 20 is shown only as the per-groupα g = .05sensitivity threshold. Routing uses the first crossing, not the terminal or maximum value. reflect both intervention advantageq g and the frequencyd g of informative comparisons. This quantity describes potential monitoring efficiency rather than realized compute savings. A hash of a fixed seed and example identifier fixes order before outcomes are examined. Each group follows the same auditable sequence: generate paired predictions, update only at discordances, selectTat the first crossing, and otherwise retainD. The resulting route, threshold, group set, models, and backends are frozen before test evaluation; test outcomes cannot revise them. 3 Empirical evaluation 3.1 Setting We use 15 languages on seven-class SIB-200, with 800 discovery and 204 test examples per language, and 15 locales on 60-class MASSIVE, with 2,033 discovery and 2,974 test examples per locale. Each dataset has 14 translation-eligible non-English groups; English is fixed to the direct path. The direct path uses multilingual MiniLM; the intervention uses pinned OPUS-MT or NLLB translation and English MiniLM. Neither classifier is task-fine-tuned. We compare direct-only, translation-only, the fixed tier of Ben Saad and Madiouni (2026), and an exact one-sided paired-test router. Accuracy is the inferential target; Macro-F1, translation rate, and latency describe the resulting policy. Full configurations, per-language effects, confidence intervals, and backend details are deferred to the supplement. Conditional on the frozen routing decisions, policy-level accuracy intervals use 2,000 paired bootstrap replicates with test examples resampled within language. Code, configurations, and reproduction scripts are available in the public repository. 1 3.2 Operating-characteristic analysis Separately from model evaluation, we simulate Bernoulli discordance streams at null and alternative values ofq g . We use 10,000 repetitions, monitoring budgets of 50, 100, 250, and 500 discordances, using the primary threshold 280 and per-group sensitivity threshold 20. We also simulate 5,000 routers under independent null streams using each datasetâs observed group-specific discordance horizons. These analyses use no test outcomes or additional model inference. 3.3 Discovery evidence SIB-200 selects Swahili, Bengali, Telugu, and Amharic. At the primary threshold 280, their first crossings occur after 23, 66, 23, and 16 discordances, respectively, and the same four routes are already selected at a budget of 200 examples per language. Afrikaans is the closest rejection: translation gains 1.13 discovery accuracy points and its running maximum reaches 18.51, below both thresholds. MASSIVE selects all 14 eligible locales, with primary-threshold crossings from 11 discordances for several locales to 141 for Polish (Figure 1). This contrast demonstrates task-dependent routing in the two evaluated settings: the evidence boundary belongs to the task and inference paths. 1 https://github.com/WajdiBenSaad/NeurIPS2026_E-Values 3 Table 1: Monte Carlo null crossing rates (%). Threshold 280 is the primary familywise-controlled decision threshold; threshold 20 is the per-groupα g = .05sensitivity threshold. The per-stream column uses a 500-discordance horizon; router columns use 14 independent null streams and observed group-specific horizons. RuleThreshold Per stream SIB router MASSIVE router Primary familywise α g = .05/142800.262.843.36 Per-group sensitivity α g = .05203.3437.3238.98 3.4 Calibration, power, and evidence requirements At the boundary nullq = .5, the empirical probability of crossing the primary threshold 280 by 500 discordances is 0.26%. Power rises from 9.50% atq = .55to 78.76% atq = .60and 99.87% at q = .65. At the per-group sensitivity threshold 20, corresponding power is 33.33%, 94.84%, and effectively one, with a boundary-null crossing rate of 3.34% (95% CI: 3.01â3.71%). Thus familywise control requires more evidence, while power remains high atq = .60. Table 1 reports the null operating points; the full power curves are in Appendix A. 3.5 Held-out confirmation All 28 frozen decisions agree with the direction of the held-out confirmatory test accuracy effect. On SIB-200, translation raises accuracy by 23.04â42.16 points for the four selected languages and has a negative held-out point estimate for every rejected non-English language. The e-router therefore translates 26.67% of examples and obtains 74.05% accuracy, an 8.14-point gain over direct-only (95% CI: 6.96â9.25), and slightly above the fixed tier while translating one fewer language. On MASSIVE, translation improves test accuracy for every eligible locale by 4.94â42.77 points. The e-router coincides with translation-only, consistent with positive held-out accuracy point estimates for all eligible locales. Relative to direct-only, its accuracy gain is 16.70 points (95% CI: 16.28â17.14). This provides no translation saving, but confirms that broad intervention is supported for that task. On SIB-200, selective routing reduces mean latency from 1.176 s for translation-only to 0.347 s per example (70.5% lower); on MASSIVE, it matches translation-only at 0.372 s because all eligible locales are translated. 3.6 Robustness and interpretation Importantly, all observed routing decisions are unchanged when moving from the per-group threshold 20 to the familywise-controlled threshold 280. Thus, the stronger router-wide error guarantee does not alter the empirical policy selected on either dataset. The primary rule yields modeled router-wide null crossing rates of 2.84â3.36%, compared with 37.32â38.98% in the per-group sensitivity analysis. Routes are also unchanged forαâ.01,.025,.05and across 50 outcome-independent orderings, including the primary ordering. The exact paired-test router selects the same groups; the contribution is therefore not superior endpoint selection, but valid continuous monitoring, first-crossing evidence, and conversion of that evidence into a reproducible frozen policy. 4 Conclusion and limitations The present process targets a positive paired-accuracy effect for each groupâtask population; it does not yet encode a practical improvement margin or intervention cost. The measured effects belong to the complete translation-and-classification path. Selective translation improves computational efficiency but may reinforce performance disparities across linguistic groups. Within these limits, paired e-processes replace fixed group heuristics with auditable, task-specific routing decisions while keeping statistical evidence distinct from downstream utility. 4 References David Ifeoluwa Adelani et al. SIB-200: A simple, inclusive, and big evaluation dataset for topic classification in 200+ languages and dialects. In EACL, 2024. Mikel Artetxe et al. Revisiting machine translation for cross-lingual classification. In EMNLP, 2023. Wajdi Ben Saad and Safa Madiouni. A cost-efficient routing pipeline for multilingual short-text classification using small language models. arXiv:2608.10939, 2026. Lingjiao Chen et al. FrugalML: How to use ML prediction APIs more accurately and cheaply. In NeurIPS, 2020. Yo Joong Choe and Aaditya Ramdas. Comparing sequential forecasters. Operations Research, 72(4):1368â1387, 2023. Jack FitzGerald et al. MASSIVE: A 1M-example multilingual natural language understanding dataset with 51 typologically-diverse languages. In ACL, 2023. Peter GrĂŒnwald et al. Safe testing. Journal of the Royal Statistical Society: Series B, 86(5):1091â1128, 2024. Anany Kotawala. Resolution diagnostics for paired LLM evaluation. ICML Workshop on Hypothesis Testing; arXiv:2605.30315, 2026. Yingfeng Luo et al. RouteLMT: Learned sample routing for hybrid LLM translation deployment. In ACL Industry Track, 2026. Quinn McNemar. Note on the sampling error of the difference between correlated proportions or percentages. Psychometrika, 12:153â157, 1947. Isaac Ong et al. RouteLLM: Learning to route LLMs from preference data. In ICLR, 2025. Aaditya Ramdas et al. Game-theoretic statistics and safe anytime-valid inference. Statistical Science, 38(4):576â 601, 2023. Zayx Shawn. PACE: Anytime-valid acceptance tests for self-evolving agents. arXiv:2606.08106, 2026. Vladimir Vovk and Ruodu Wang. E-values: Calibration, combination and applications. The Annals of Statistics, 49(3):1736â1754, 2021. Ian Waudby-Smith and Aaditya Ramdas. Confidence sequences for sampling without replacement. In NeurIPS, 2020. Wei-Chi Wu et al. No one fits all: From fixed prompting to learned routing in multilingual LLMs. In Findings of ACL, 2026. A Technical appendices and supplementary material The supplement contains frozen policy results, detailed latency summaries, policy-level paired- bootstrap accuracy intervals, a synthetic threshold-crossing power curve, a confirmatory qualityâ translation trade-off, and discovery-budget sensitivity; the NeurIPS paper checklist follows. These materials use validated stored predictions or synthetic discordance streams; no additional model inference or translation was performed. Compute resources. All experiments ran locally on a single MacBook Pro (MacBookPro16,1) with a 2.6 GHz six-core Intel Core i7, 16 GB RAM, and an AMD Radeon Pro 5300M GPU with 4 GB VRAM. The automatic device setting selected PyTorch MPS acceleration. Encoder and translation batch sizes were 128 and 8, respectively, with one data worker; languages were processed sequentially. Summed per-language elapsed times were 3.39 h and 1.09 h for SIB-200 discovery and evaluation, and 2.83 h and 4.72 h for MASSIVE discovery and evaluation, totaling 12.04 h for the final runs. These figures exclude manual pauses between language launches and exploratory or smoke-test runs. Subsequent statistical analyses reused cached predictions and required no model inference or translation. 5 Table 2: Frozen application-level policy results. Translation and accuracy are percentages. DatasetPolicyAccuracyMacro-F1Translation SIB-200Direct65.92.66550.00 Fixed tier73.82.738733.33 E-router74.05.740726.67 Translate69.25.696393.33 MASSIVEDirect27.36.30410.00 Fixed tier39.30.362733.33 E-router44.07.369493.33 Translate44.07.369493.33 Table 3: Measured policy latency in seconds per example. Times are assembled from steady-state complete-path measurements and exclude model loading and cold starts; they are hardware-specific. DatasetPolicyMeanMedianP95 SIB-200Direct.005.005.008 Fixed tier.461.0051.710 Paired p-value.347.0051.691 E-router.347.0051.691 Translate1.1761.1901.988 MASSIVEDirect.002.002.002 Fixed tier.153.002.592 Paired p-value.372.382.592 E-router.372.382.592 Translate.372.382.592 Table 4: Frozen e-router accuracy gain over direct-only on the held-out confirmatory test sets. Percentile 95% confidence intervals use 2,000 paired bootstrap replicates with examples resampled within language. DatasetNGain (points)95% CI SIB-2003,0608.14[6.96, 9.25] MASSIVE44,61016.70[16.28, 17.14] 0.30.40.50.60.70.80.9 Discordant win probability q 0.0 0.2 0.4 0.6 0.8 1.0 Rejection / crossing probability Primary routing threshold: 280 Primary e-process (threshold 280) Per-group sensitivity (threshold 20) Fixed-horizon exact test = 0.05 Figure 2: Probability of an e-process threshold crossing by 500 discordances under Bernoulli alternatives. The primary routing curve uses the familywise-controlled threshold 280; threshold 20 is shown only as a per-group sensitivity analysis. The fixed-horizon exact test is an endpoint reference. 6 020406080 Translation rate (%) 0.675 0.700 0.725 0.750 Accuracy Direct only Fixed list Paired p-value / E-value Translation only SIB-200: Accuracy 020406080 Translation rate (%) 0.68 0.70 0.72 0.74 Macro-F1 Direct only Fixed list Paired p-value / E-value Translation only SIB-200: Macro-F1 020406080 Translation rate (%) 0.30 0.35 0.40 0.45 Accuracy Direct only Fixed list Paired p-value / E-value / Translation only MASSIVE: Accuracy 020406080 Translation rate (%) 0.32 0.34 0.36 0.38 Macro-F1 Direct only Fixed list Paired p-value / E-value / Translation only MASSIVE: Macro-F1 Figure 3: Confirmatory qualityâtranslation trade-off. Coincident points have identical frozen routes. 7 NeurIPS Paper Checklist 1. Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paperâs contributions and scope? Answer: [Yes] Justification: The abstract and introduction state the paperâs methodological contribution, statistical guarantees, empirical scope, and task-specific conclusions consistently with Sections 2â4. 2. Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: Section 4 discusses the current accuracy-based target, lack of explicit util- ity/cost modeling, and dependence on the complete translation-and-classification pipeline. 3. Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [Yes] Justification: Sections 2.2â2.3 state the conditional-null assumptions, define the filtration, justify the e-process property, and derive the anytime-valid crossing guarantee via the e-process maximal inequality. 4. Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: The main paper specifies datasets, splits, model families, routing rules, statisti- cal thresholds, simulation settings, and evaluation metrics; exact configurations and backend details are provided in the supplementary material and code. 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: A public code package accompanies the submission and includes environment specifications, configuration files, scripts, and instructions for reproducing the reported experiments from the public datasets. 6. Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyperpa- rameters, how they were chosen, type of optimizer) necessary to understand the results? Answer: [Yes] Justification: Section 3 specifies the datasets, discovery/test sizes, routing baselines, model families, statistical thresholds, and evaluation protocol; exact model revisions and translation backends are documented in the supplement. 7. Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] 8 Justification: Policy-level accuracy gains are reported with percentile 95% confidence inter- vals from 2,000 paired bootstrap replicates with test examples resampled within language; the statistical route-discovery procedure is also accompanied by null-crossing and power simulations. 8. Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: The supplement reports the local hardware, accelerator, memory, batching and worker settings, execution strategy, and recorded wall-clock time for each final discovery and evaluation run; it also distinguishes these runs from exploratory analyses and manual pauses. 9. Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.c/public/EthicsGuidelines? Answer: [Yes] Justification: The work uses established public benchmark datasets and publicly available pretrained models and does not involve human-subject data collection, deceptive experimen- tation, or restricted personal data. 10. Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: Section 4 discusses the computational-efficiency benefits of selective transla- tion as well as the risk of reinforcing performance disparities across linguistic groups. 11. Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)? Answer: [N/A] Justification: The work does not release a new high-risk model, scraped sensitive dataset, or other asset requiring special misuse safeguards. 12. Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: All datasets and pretrained models are credited to their original sources, and their versions, licenses, and terms of use are documented in the supplementary material and code package. 13. New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: The released experiment code is documented with setup instructions, configu- ration files, reproducibility commands, and descriptions of generated artifacts. 14. Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? 9 Answer: [N/A] Justification: The work uses existing benchmark datasets and does not conduct crowdsourc- ing or experiments involving human participants. 15.Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? Answer: [N/A] Justification: The study does not involve human-subject recruitment, intervention, or collec- tion of new human-subject data. 16. Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required. Answer: [N/A] Justification: LLMs are not an important or non-standard component of the core method- ology; the study evaluates pretrained multilingual and English encoders and machine- translation models. 10