WorksComputer Science
From Unverifiable Domain Text to Cross-Lingual RLVR via Code-Switched Task Synthesis
Abstract
Reinforcement Learning with Verifiable Rewards (RLVR) can improve reasoning in large language models, but specialized non-English domains expose a data mismatch. Raw English domain text is often available, while target-language RLVR tasks with verifiable rewards are scarce. We introduce code-switched task synthesis for cross-lingual RLVR, a data-centric approach that derives target-language and code-switched RLVR task variants from otherwise unverifiable English domain text. Using fill-in-the-middle multiple-choice QA as a seed task format, we synthesize automatically verifiable tasks from domain passages and reshape them into cross-lingual variants. These variants preserve the same evidence-answer relation but change how English and target-language evidence are presented during RLVR. On Korean medical benchmarks and cybersecurity benchmarks across four target languages, code-switched training outperforms matched monolingual and bilingual-mixture baselines on average target-language performance while maintaining comparable English performance. Additional analyses provide evidence of better target-language input understanding and stronger cross-lingual alignment in both model outputs and internal representations. These results suggest that code-switched training helps target-language inputs access domain reasoning learned more strongly in English.
Introduction
Reinforcement Learning with Verifiable Rewards (RLVR) has become a powerful post-training method for improving the reasoning ability of large language models ([15] (Shao et al. 2024)). Its effectiveness depends on tasks whose
outcomes can be automatically verified and converted into reward signals. Mathematics and programming naturally provide such tasks through exact answers, symbolic checks, or execution, but the same assumption does not generally hold for knowledge-intensive domains. Although raw text is abundant, scaling the creation of high-quality verifiable problems is difficult and time-consuming. We study this setting using cybersecurity and medicine as controlled domain testbeds (Zhang et al., 2025; Su et al., 2026). This limitation is more severe outside English, where domain-specific multilingual corpora are scarce and rewardable reasoning data is even scarcer ([8] (Wang et al. 2025b)). Thus, a central obstacle for cross-lingual domain RLVR is not only how to obtain multilingual data, but how to turn that data into verifiable training signals.
A practical way to construct verifiable supervision from domain text is fill-in-the-middle multiple-choice question answering (FIM-MCQA) synthesis ([26] (Ren et al. 2025)). FIM-MCQA masks a span from a domain passage, treats the removed span as the correct answer, and rewards the model for selecting it. This turns otherwise unverifiable raw text into an automatically verifiable task. In multilingual settings, the verified seed task can be translated into a target language to construct target-language training data.
Yet translation alone may not provide the most effective transfer signal. Multilingual reasoning gaps often stem from input understanding or cross-lingual alignment failures ([20] (Son et al. 2026)), and specialized domains may require target-language inputs to access stronger English-side domain knowledge. Code-switching offers such a bridge by placing English and target-language evidence in the same context, but it has mainly been studied in pretraining, supervised fine-tuning, or inference-time settings (Chen et al., 2024; Zhu et al., 2024; Etxaniz et al., 2024; Yoo et al., 2026; Wang et al., 2025c; Yoo et al., 2025). We instead ask how to make code-switching part of the verifiable RLVR training signal itself.
We introduce code-switched task synthesis for cross-lingual RLVR as a data-construction approach for specialized domains where English text is available but target-language RLVR data with verifiable rewards is scarce. We synthesize FIM-MCQA seed tasks from cybersecurity and medical corpora and derive target-language and code-switched training variants that preserve the same evidence-answer relation. This design keeps the training signal compatible with standard answer-based RLVR while changing how the domain evidence is presented across languages. We evaluate the resulting data on Korean medical and cybersecurity benchmarks and in Swahili, Hindi, and Japanese cybersecurity transfer settings. The code-switched data improves target-language performance over monolingual and bilingual-mixture baselines, with comparable English performance. Input-level, representation-level, and reasoning-trace analyses suggest that code-switched training helps models leverage English-dominant domain reasoning for target-language beyond translated and bilingual variants.
Contributions. We introduce code-switched task synthesis for cross-lingual RLVR as a data-construction approach for domains where English text is available but target-language RLVR data with verifiable rewards is scarce. Starting from English FIM-MCQA seed tasks, we derive target-language and code-switched training variants that preserve the same evidence-answer relation and remain compatible with standard answer-based RLVR. Under matched supervision, code-switched data improves target-language performance over English-only, target-language-only, and bilingual-mixture training while maintaining comparable English performance. Analyses of answer consistency, input understanding, representations, and generated reasoning suggest that code-switched training helps target-language inputs access English-side domain reasoning beyond translated and bilingual variants.
Related Work
Verifiable Data Construction for RLVR
Reinforcement learning with verifiable rewards (RLVR) has become an effective post-training paradigm for improving reasoning [15] (Shao et al. 2024). Math and programming tasks naturally support RLVR because their answers can often be checked directly, for example by exact matching or execution. In specialized, data-scarce domains, the available data often consists of raw documents rather than rewardable MCQA examples. Building verifiable MCQA datasets by hand is costly and time-consuming.
Recent work addresses this challenge by constructing automatically verifiable tasks from text. General-Reasoner mines and re-verifies web QA pairs [29], Nemotron-CrossThink aggregates multi-domain reasoning corpora [1], and Golden Goose synthesizes fill-in-the-middle multiple-choice QA (FIM-MCQA) tasks from otherwise unverifiable text [26, 16]. Domain-specific resources also exist in cybersecurity or biomedical corpora mainly designed for supervised training (Yu et al., 2025b; Touchent et al., 2026). However, prior work does not directly address how to construct verifiable domain tasks into cross-lingual training signals when target-language data is scarce from the raw text.
Our work follows the FIM-MCQA direction but uses it for a different purpose. A translated FIM-MCQA task gives target-language exposure, but it still keeps English and the target language as separate versions of the same content. We instead construct code-switched FIM-MCQA tasks, where both languages appear within a single rewardable instance. This lets the model learn the cross-lingual bridge through the same answer reward.
Multilingual Reasoning Gaps and Code-Switched Transfer
Large language reasoning models exhibit substantial gaps between English and other languages, especially in low-resource settings [8, 28, 21] (Wang et al. 2025b; Shi et al. 2023). Recent analyses attribute much of this gap to input understanding, where the model fails to map non-English inputs into its dominant reasoning language [20]. This view is consistent with work showing that hub-language reasoning often remains more accurate than forced target-language reasoning (Tan et al., 2025; Qi et al., 2025), and with mechanistic analyses identifying dominant languages as privileged reasoning channels in multilingual LLMs (Schut et al., 2025; Zhong et al., 2024; Wang et al., 2025a). Taken together, these findings suggest that a central bottleneck is how target-language inputs can access English-side domain knowledge and reasoning behavior.
Prior work has addressed this bottleneck through both algorithmic and data-centric routes. Algorithmic approaches modify the training framework, reasoning language, or reward design, especially in math reasoning and machine translation settings where automatic rewards are easier to define [11] (Zhang et al. 2026). Data-centric approaches instead change the supervision or prompting interface used for transfer, including translated or question-aligned SFT [6] (Zhu et al. 2024), cross-lingual alignment or bridging methods (She et al., 2024; Yoon et al., 2024), and inference-time translation (Etxaniz et al., 2024). These methods show that language form can strongly affect cross-lingual reasoning, but they do not study how mixed-language data should be synthesized for RLVR training.
Code-switching, alternating between languages within a single context [3], provides a natural data-centric interface for this problem because it places source-language and target-language evidence in the same instance. Prior work uses code-switching as a cross-lingual transfer signal in pre-training with synthetic data (Wang et al., 2025c), supervised fine-tuning with curriculum learning or chain-of-thought supervision (Yoo et al., 2025; Son et al., 2026), and inference-time demonstrations (Yoo et al., 2026). We extend this line to multilingual RLVR by synthesizing code-switched FIM-MCQA tasks from English domain corpora. The resulting tasks expose the model to English and target-language evidence within the same rewardable instance, so cross-lingual transfer is learned from verifiable multilingual training data rather than from translation-only supervision. To our knowledge, this is the first data-centric approach that uses code-switched RLVR tasks to induce cross-lingual transfer in multilingual settings.
Method
Overview
We convert raw English domain corpora into cross-lingual RLVR training tasks for a resource-mismatch setting in which English domain text is available but target-language verifier-ready RLVR data is scarce. We build a source pool from cybersecurity (Yu et al., 2025b) and medical corpora (Touchent et al., 2026). We remove boilerplate, non-educational pages, and low-quality documents. Appendix § 9.§ 1 reports the filtering rules.
Figure 1 illustrates the pipeline. First, we synthesize English FIM-MCQA seed tasks from the filtered documents. Each task marks a verbatim span in a self-contained passage and uses the removed span as the answer. Second, we derive target-language and code-switched variants from each seed task. The variants keep the same domain content but change its language form, allowing us to compare translation-only exposure, option-side code-switching, and passage-level code-switching under matched supervision.

Figure 1. Data construction pipeline for cross-lingual RLVR from raw English domain corpora. (a) We first extract educational passages from domain raw corpora, identify answer spans, and generate distractors to synthesize English FIM-MCQA seed tasks. We then translate each seed task into the target language and construct code-switched variants: option-side code-switching, which pairs passages and answer options across languages, and passage-level code-switching, which mixes English and target-language evidence within the passage in both directions. (b) The resulting training variants used for controlled RLVR comparisons.
Verifiable FIM-MCQA Synthesis
For each retained source document , we synthesize an English FIM-MCQA seed task , where is a passage with one span replaced by [MASK], is the answer-option set, and is the correct option index. The central constraint is that the correct answer must originate from a verbatim span in the passage. This keeps the task compatible with an automatic answer-option verifier during RLVR.
Rather than asking an LLM to generate the full task in one call, we decompose synthesis into extraction, span verification, and distractor generation. The synthesizer first extracts a self-contained educational passage from the filtered source document. It then selects a consecutive answer span from and is instructed to copy the span verbatim. A deterministic verifier checks that appears in and replaces that occurrence with [MASK] to form . Examples that fail this substring check are retried or discarded.
Finally, the model generates plausible distractors conditioned on . The distractors are required to match the length, format, and semantic type of the answer while avoiding direct answer leakage. The correct span and distractors are shuffled to form , with the correct index recorded as . This decomposition prevents paraphrased or non-substring answers from entering the pool while keeping the reward function simple and fully automatic. The complete multi-stage extraction pipeline, and prompt templates are provided in Appendix § 9.§ 2.
Cross-Lingual Task Variant Synthesis
Given an English FIM-MCQA seed task, we construct target-language and code-switched variants by changing how the same domain content is presented across languages. We first translate the masked passage and answer options into the target language, yielding , with the [MASK] marker preserved.
Translation gives target-language exposure, but it keeps English and the target language in separate examples. We therefore construct code-switched variants that bring English and target-language evidence into the training task, making language form the main intervention.
Let , and write for the masked passage in language . For and , we use two code-switched variants that differ in where the language boundary is placed.
Option-side code-switching. We pair a passage in one language with answer options in the other:
This trains cross-lingual matching between passage evidence and answer choices.
Passage-level code-switching. We mix languages inside the passage while keeping the options in the base language:
Here, replaces selected content-bearing spans in with fluent translations in language .
We generate both directions, and . The first introduces target-language spans into English domain passages, while the second starts from target-language passages and inserts English spans. This design targets cross-lingual transfer: it trains target-language inputs to connect with English-dominant domain evidence.
Generated passages are filtered for [MASK] preservation, target-language coverage, script validity, and malformed switching patterns. The exact prompt templates, generation configurations, and detailed filtering rules are reported in Appendix § 9.§ 3.
Experiments
Experimental Setting
Experimental design. We design the experiments to test whether code-switched RLVR improves cross-lingual transfer in specialized domains where English resources are abundant but target-language verifier-ready RLVR data is scarce. We use medicine and cybersecurity as two controlled domain settings. Korean is our main target language because it provides native medical benchmarks while remaining linguistically distant from English. We additionally evaluate Swahili, Hindi, and Japanese in the cybersecurity domain to broaden language and script coverage. English serves as the source language for FIM-MCQA synthesis and for constructing target-language and code-switched views.
Base models. We use three 3B-scale open-source language models: Llama 3.2-3B-Instruct ([14]), Qwen2.5-3B-Instruct ((Yang et al., 2024)), and SmolLM3-3B ([4]). These models provide a suitable base model for studying cross-lingual RLVR. These models can process target-language inputs, but their domain-level reasoning behavior is still largely English-dominant. This setting allows us to test whether target-language inputs can better leverage reasoning behavior learned most strongly in English through various settings.
Training variants and RL setup. We compare the original checkpoint (BASE) with several RL-trained variants. RLVR-BASE is trained only on an auxiliary general-reasoning RLVR mixture, while the domain variants additionally use the synthesized domain FIM-MCQA pool from Appendix § 9.§ 4. These domain variants differ only in how each domain example is realized across languages: EN-ONLY uses English examples, TGT-ONLY uses target-language examples, BI-MONO mixes English and target-language monolingual views, and OURS combines monolingual, Option-side, and passage-level code-switched datasets while preserving the same answer index and verifier. We mix domain FIM-MCQA data with the auxiliary RLVR mixture to prevent the model from overfitting to the fill-in-the-middle MCQA format and to mitigate format collapse during RLVR training. The auxiliary mixture contains general reasoning data, such as scientific reasoning, math, coding, and instruction-following prompts ([1, 27, 5]), which acts as a regularizer for retaining broad free-form reasoning ability while the model acquires cross-lingual, domain-specific reasoning skills. Empirically, we found this mixing strategy to yield more stable RLVR training than using the domain-specific FIM-MCQA pool alone.
All RL variants are trained with DAPO ((Yu et al., 2025a)). We sample 16 rollouts per prompt, use a learning rate of , and train for two epochs. Each domain variant uses the same auxiliary sampling ratio. Additional details on the variants and training configurations are provided in Appendix § 10.§ 1.
Evaluation. We evaluate target-language and English performance in both domains. For medicine, we use MedQA ([19]), MedMCQA ((Pal et al., 2022)), KorMedMCQA-Doctor ([22]), and KMMLU-Pro-Medical ([17]). For cybersecurity, we use CTI-MCQ, CyberMetric, and SecEval ([2, 23] (Tihanyi et al. 2024)). We evaluate target-language versions of the cybersecurity benchmarks in Korean, Swahili, Hindi, and Japanese. Because this task primarily targets non-English performance, we select the best checkpoint based on target-language validation scores. The translation process and subsequent quality evaluations of translated benchmarks are described in Appendices § 10.§ 2 and § 10.§ 3, respectively.
Experimental Results
Table 1 shows the medical results for Qwen2.5-3B-Instruct and Llama-3.2-3B-Instruct. RLVR improves performance across languages, regardless of whether the domain examples are presented in English, Korean, or mixed-language form. The gains are particularly clear when the training mixture includes in-domain medical FIM-MCQA data. EN-ONLY and TGT-ONLY often outperform RLVR-BASE, showing that the improvement comes not only from additional RL optimization but also from rewardable examples in the target domain.
We next compare different language realizations of the same domain data. BI-MONO exposes the model to both English and Korean, but it does not consistently outperform the monolingual domain baselines. This suggests that simply adding translated views or mixing monolingual examples across languages is not enough to reliably improve targetlanguage reasoning.
The main improvement appears when code-switched examples are added. Across both medical model families, OURS gives more consistent gains on Korean benchmarks than the monolingual and bilingual baselines. On Korean-native benchmarks such as KMMLU-Pro-Medical and KorMedMCQA-Doctor, it improves over all non-code-switched baselines by more than one point in several settings. It also improves translated medical QA performance over both EN-ONLY and TGT-ONLY, indicating that code-switching helps not only on native Korean benchmarks but also on Korean versions of English medical reasoning tasks.
A similar pattern appears beyond this Korean medical setting. As presented in Table 2, cybersecurity results across four target languages show that code-switched RLVR again yields stronger average target-language performance than monolingual or bilingual-mixture training. English performance, reported as Avg (En), remains comparable across variants. These results broaden the language coverage of the main finding, although the size of the improvement varies across individual benchmarks, models, and domains. Overall, the results support code-switched training as an effective cross-lingual transfer signal that helps the model apply domain reasoning learned more strongly in English to target-language inputs. Full results, including base-model performance and per-benchmark English scores, are provided in Appendix § 11.§ 1. We also report the multi-seed experiment on the Korean cybersecurity setting in Appendix § 11.§ 2.
| Bench. | Korean: RLVR-Base | +En | +Tgt | +BiMono | +Ours | Swahili: RLVR-Base | +En | +Tgt | +BiMono | +Ours |
| SecEval* | 44.3 | 50.1 | 49.8 | 51.1 | 57.3 | 25.8 | 26.8 | 28.6 | 25.9 | 28.3 |
| CyberMetric* | 65.8 | 70.6 | 80.2 | 72.4 | 80.0 | 54.0 | 53.9 | 54.2 | 54.5 | 55.0 |
| CTI-MCQ* | 40.6 | 43.2 | 45.5 | 43.8 | 47.1 | 41.9 | 42.1 | 43.2 | 43.4 | 45.0 |
| Avg (Tgt) | 50.23 | 54.63 | 58.50 | 55.77 | 61.47 | 40.57 | 40.93 | 42.00 | 41.27 | 42.77 |
| Bench. | Hindi: RLVR-Base | +En | +Tgt | +BiMono | +Ours | Japanese: RLVR-Base | +En | +Tgt | +BiMono | +Ours |
| SecEval* | 39.2 | 54.6 | 54.9 | 53.1 | 56.0 | 50.6 | 51.3 | 52.2 | 52.0 | 53.4 |
| CyberMetric* | 66.7 | 71.2 | 73.5 | 74.5 | 77.0 | 77.4 | 77.4 | 78.4 | 77.8 | 79.8 |
| CTI-MCQ* | 48.0 | 50.9 | 51.4 | 50.7 | 51.4 | 49.7 | 51.1 | 51.2 | 52.0 | 52.3 |
| Avg (Tgt) | 51.30 | 58.90 | 59.93 | 59.43 | 61.47 | 59.23 | 59.93 | 60.60 | 60.60 | 61.83 |
Table 2. Comprehensive performance on cybersecurity benchmarks across RLVR data configurations for four target languages spanning different scripts and resource levels: Korean and Swahili (top), and Hindi and Japanese (bottom). All other notations and formatting follow Table 1.
| Bench. | Qwen2.5-3B-Instruct | Llama-3.2-3B-Instruct | ||||||||
| RLVR-Base | +En | +Tgt | +BiMono | +Ours | RLVR-Base | +En | +Tgt | +BiMono | +Ours | |
| KorMedMQ-D | 35.06 | 35.34 | 35.29 | 35.34 | 36.67 | 21.10 | 27.93 | 27.47 | 28.79 | 29.31 |
| MedMCQA* | 34.52 | 33.96 | 34.03 | 33.72 | 34.86 | 24.26 | 27.12 | 27.36 | 29.65 | 29.22 |
| KMMLU-Pro-M | 35.45 | 35.76 | 36.21 | 36.06 | 36.29 | 22.84 | 24.77 | 24.09 | 25.08 | 26.36 |
| Avg (Tgt) | 35.01 | 35.02 | 35.18 | 35.04 | 35.94 | 22.73 | 26.61 | 26.31 | 27.84 | 28.30 |
| Avg (En) | 51.42 | 52.19 | 52.97 | 52.96 | 52.13 | 51.65 | 56.64 | 56.89 | 57.21 | 57.26 |
Table 1. Performance on medical benchmarks across RLVR data configurations. RLVR-BASE denotes RLVR training using only the auxiliary data pool. The remaining variants include the same auxiliary pool and additionally use English data (+EN), target-language data (+TGT), bilingual monolingual data (+BIMONO), or our proposed code-switched data (+OURS). The asterisk (*) indicates that the benchmark is translated into the respective language. The gray Avg (En) row reports average English performance (details in Appendix § 11.§ 1). Bold indicates the best result within each model, and green shading highlights our proposed method (+OURS).
Component Ablation
To examine how each form of code-switching contributes to the target-language improvement, we conduct a component ablation in the Korean cybersecurity setting with SmolLM3-3B. All ablation variants follow the same experimental setup and differ only in the included code-switching component. OPT-ONLY augments BI-MONO with option-side code-switching, PASS-ONLY augments it with passage-level code-switching, and OURS combines both components.
Table 3 shows that both code-switching components improve the target-language average over BI-MONO. Passage-level switching yields a larger improvement than option-side switching, while combining both components produces the strongest target-language average and outperforms either component alone on all three target-language benchmarks. English performance remains comparable across the variants. These results suggest that passage-level switching contributes more strongly in this setting, while option-side switching provides an additional benefit when combined with passage-level switching. Detailed results are provided in Appendix § 11.§ 3.
| Variant | Target Avg | English Avg |
| RLVR-Base | 50.23 | 64.87 |
| + En | 54.63 | 67.00 |
| + Tgt | 58.50 | 67.83 |
| + BiMono | 55.77 | 66.80 |
| + Opt | 56.30 | 67.20 |
| + Pass | 59.37 | 67.07 |
| + Ours | 61.47 | 68.27 |
Table 3. Component ablation on Korean cybersecurity benchmarks (SmolLM3-3B). Values report the macroaverage accuracy across benchmarks in the Target Language (Korean) and English.
Cross-lingual Consistency
To test whether code-switched RLVR reduces the language barrier between English and Korean inputs, we measure accuracy-preserving cross-lingual consistency on parallel cybersecurity questions. For each English–Korean question pair, we partition model predictions into three mutually exclusive cases: both languages correct, English-only correct and Korean-only correct. The both-correct rate captures language-invariant correct decisions, whereas the English-only and Korean-only gap captures directional answer asymmetry.
As shown in Table 4, OURS achieves the strongest accuracy-preserving consistency and reduces asymmetry. This suggests that the code-switched data does more than add translated supervision: it helps Korean input elicit the same answer decisions as their English counterparts, providing empirical evidence for the cross-lingual bridge.
| (English, Korean) | (✓,✓) ↑ | (✓,×) ↓ | (×,✓) ↓ |
| Base | 15.1 | 28.2 | 13.0 |
| En-only | 28.1 | 30.0 | 7.3 |
| Bi-Mono | 29.2 | 29.8 | 7.1 |
| Ours | 34.8 | 27.1 | 6.7 |
Table 4. Cross-lingual consistency comparison on cybersecurity benchmarks. This table aggregates the outcomes of English and Korean question pairs across cybersecurity domain benchmarks. ✓ denotes a correct answer, and × denotes an incorrect answer. Bold indicates the most consistent results.
Analysis
Table 4 shows that Code-switched RLVR improves cross-lingual consistency, but the source of this gain is not obvious from answer accuracy alone. Following the view that multilingual reasoning gaps arise when target-language inputs fail to reach the model’s dominant, often English-proximal, reasoning pathway, we test whether Ours reduces this gap at three levels: input understanding, internal representation, and generated reasoning behavior.
In-domain Translation Quality
We first test whether the Ours gain begins at the input-understanding stage. Prior work argues that multilingual reasoning gaps often arise because the model fails to map non-English inputs into the dominant language of its reasoning trace, typically English [20]. We therefore use in-domain Korean-to-English translation quality as a diagnostic proxy for how well a model recovers the English-side semantic content of Korean questions. We compare this score with the reasoning performance ratio, defined as Korean accuracy divided by accuracy on the matched English benchmark. Details of the translation prompt, GEMBA-DA scoring, and metrics are provided in Appendix § 12.§ 2.
Figure 2 shows that in-domain translation quality closely tracks retained Korean reasoning performance. The correlation is near-perfect on Cyber-Metric () and remains high on KMMLU-Pro-Medical (). In both domains, the code-switched model lies toward the upper-right region, indicating that it both recovers the English-side meaning of Korean inputs more accurately and preserves a larger fraction of its English reasoning performance. This supports the view that code-switched RLVR reduces the target-language understanding bottleneck before reasoning begins.
Figure 2. Reasoning performance ratio vs. in-domain Korean-to-English translation quality for (a) CyberMetric (SmolLM3-3B) and (b) KMMLU-Pro-Medical (Llama-3.2-3B-Instruct). Dashed lines show linear trends; denotes Pearson correlation.
Latent Reasoning State Alignment
The translation analysis suggests that part of the OURS gain comes from better semantic recovery of Korean inputs. We next ask whether this gain is visible at the point where the model begins to reason. We analyze the SmolLM3-3B ablation models on a fixed parallel English–Korean subset of CyberMetric. For each question pair, we append the same opening reasoning token, <think>, and extract the hidden state at that token before any reasoning chain-of-thoughts(CoTs) are generated. We measure English–Korean alignment across layers with Mean-Centered Cosine and Average Neuronwise Correlation (ANC) (Del and Fishel, [7]). Details of experiment settings and metric definitions are provided in Appendix § 12.§ 3.
Figure 3 shows a common depth pattern across models: English–Korean similarity rises in the middle layers, where cross-lingual semantic representations are expected to become more comparable. The important difference is the magnitude of this convergence. In the highlighted layer band, OURS consistently achieves higher similarity than the non-code-switched variants across both measures. Thus, the baselines do show some middle-layer alignment, but their English and Korean reasoning-entry states remain less tightly matched than those of OURS. This is especially informative because BI-MONO already includes bilingual supervision; exposure to both languages alone does not produce the same level of alignment as adding code-
Figure 3. Layer-wise English–Korean representation alignment on CyberMetric, measured at the opening <think> marker. Our method shows stronger middle-to-upper layer alignment than the non-code-switched baselines. The gray area marks middle layers 12–24.
switched training dataset.
This depth profile is consistent with prior findings that multilingual models often form their most comparable cross-lingual semantic states in the middle layers ([30, 24]), and with the view that non-English inputs may pass through an English-proximal reasoning space before target-language realization (Schut et al., 2025). Under this interpretation, code-switched RLVR does not create cross-lingual alignment from scratch. Instead, it amplifies the natural middle-layer convergence and narrows the English–Korean gap at the reasoning-entry state.
Generated Reasoning Trace Alignment
The representation analysis above suggests that code-switched RLVR better aligns English and target-language inputs. We further examine whether this alignment is reflected in the model’s generated reasoning CoTs. To this end, we conduct a pairwise LLM-as-a-judge evaluation that compares OURS with monolingual RLVR baselines. For each example, the judge is given English and target-language outputs from two systems and selects the system whose two outputs are more semantically consistent in high-level reasoning, problem decomposition, and solution strategy. We use GPT-5.5 as the judge model and swap the order to mitigate position bias. The full evaluation details are provided in Appendix § 12.§ 4.
Table 5 shows that OURS consistently outperforms EN-ONLY across all evaluated model-domain settings, with gains of +8.3, +8.2, and +4.6 points for Qwen2.5-3B-Instruct, Llama-3.2-3B-Instruct, and SmolLM3-3B, respectively. Compared with TGT-ONLY, OURS is comparable in the medical domain and substantially better in cybersecurity. These results suggest that code-switched RLVR helps the model leverage English-side domain reasoning when solving target-language problems.
| vs. EN-ONLY | vs. TGT-ONLY | |||
| Win Rate(%) | Ours | En-only | Ours | Tgt-only |
| Medical | ||||
| Qwen2.5 | 48.3 | 40.0 | 47.5 | 42.5 |
| Llama3.2 | 47.0 | 38.8 | 43.3 | 42.0 |
| Cyber | ||||
| SmolLM3 | 36.3 | 31.7 | 58.7 | 21.7 |
Table 5. Pairwise judge win rates to measure cross-lingual reasoning consistency using GPT-5.5. The target language is Korean. The reported results are aggregated across benchmarks within each domain (evaluating 100 samples per benchmark). Our method demonstrates consistent improvement across all models and baselines.
Conclusion
This work proposes cross-lingual RLVR in specialized domains as a data construction problem to alleviate data scarcity issue in the RLVR framework. We build on FIM-MCQA synthesis to obtain rewardable seed tasks from English cybersecurity and medical documents, then reshape those tasks into target-language and code-switched training data. This lets us study how the language form of the same domain evidence affects transfer under a standard RLVR training. The results show that our synthesized code-switched data provides stronger target-language transfer than translation-only or bilingual-mixture training, while English performance remains comparable. The analyses give a consistent explanation. Code-switched training improves answer agreement across languages, supports better recovery of target-language input meaning, strengthens English–target representation alignment, and produces more consistent reasoning traces. These findings suggest that code-switched task synthesis is a practical way to help models leverage English-dominant domain reasoning for target-language inputs in data-scarce multilingual domains.
Limitations
Our study uses a controlled but computationally demanding RLVR setup. We train multiple domain–language–model variants with the same RL algorithm and data mixture, and each training run requires approximately 2,500 H100 GPU hours. This cost prevents exhaustive search over training schedules, data mixtures, rollout budgets, and model scale. Prior work suggests that longer or more carefully staged RL training [25], as well as larger rollout budgets [18], can further affect reasoning performance. We therefore interpret our results as evidence for the effectiveness of code-switched task synthesis under a practical compute budget, rather than as a fully optimized upper bound. For the same reason, we focus on 3B-scale open-source models, which make controlled comparisons feasible across domains, languages, and training variants.
A second limitation is evaluation coverage. While English domain benchmarks are widely available, comparable target-language benchmarks are scarce, especially outside high-resource languages. As a result, our cybersecurity evaluation relies partly on translated benchmarks. These translations may introduce artifacts or differ from naturally written target-language questions. We mitigate this issue with domain-aware translation prompts and translation quality checks in Appendix § 10.§ 3. We also include native Korean medical benchmarks, which provide a complementary test beyond translated evaluation.
Finally, our study covers a limited range of languages, domains, and code-switching designs. Korean is evaluated most extensively, while Swahili, Hindi, and Japanese are examined only in cybersecurity under narrower experimental coverage. Prior work shows that cross-lingual transfer can vary with the form and mixture of code-switching data (Wang et al., 2025a). Although our component ablation compares option-side and passage-level switching, broader language and domain settings and alternative switching configurations remain for future work.
Acknowledgments
We would like to thank the anonymous reviewers for their constructive feedback and suggestions. We would like to thank the anonymous reviewers for their constructive feedback and suggestions. We extend our gratitude to Gichang Lee and Nako Sung for their unwavering support. This work was partly supported by Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program (KAIST)) and by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2025-00555621).
Appendix
We provide detailed supplementary materials organized as follows:
Appendix A describes how the training data is constructed.
A.1: source corpora and the data pre-filtering process.
A.2: verifiable FIM-MCQA synthesis.
A.3: cross-lingual task variant synthesis.
A.4: final data statistics.
A.5: qualitative examples of the processed data.
Appendix B describes the experimental setup.
B.1: training details.
B.2: evaluation dataset translation.
B.3: translation quality assessment of the evaluation dataset.
Appendix C reports additional quantitative results.
C.1: full evaluation results, including the no-RLVR settings.
C.2: multi-seed robustness results.
C.3: per-benchmark component ablation results.
Appendix D details our analyses.
D.1: cross-lingual consistency comparison.
D.2: in-domain translation experiment.
D.3: representation alignment analysis.
D.4: generated reasoning trace alignment experiment.
Appendix E lists the prompts utilized in our experiments.
Data Construction
Source Corpora and Pre-filtering Details
Cybersecurity For the cybersecurity domain, we use the Primus dataset (Yu et al., 2025b), an open-source corpus released by Trend Micro for training cybersecurity large language models. It consists of two sub-corpora: Primus-Seed, a manually curated, high-quality collection of documents gathered from MITRE, Wikipedia, major security vendor websites, and expert-collected Cyber Threat Intelligence (CTI); and Primus-FineWeb, which filters cybersecurity-related texts from FineWeb using a TinyBERT classifier trained with Primus-Seed as positive samples.
We combine these two sub-corpora into Primus-Pretraining and apply an additional educational quality filter. We include only documents that score (on a scale of 0 to 5) using nemocurator-fineweb-nemotron-4-edu-classifier4 (Su et al., 2025). Although the default high-quality threshold for this classifier is 2.5, cybersecurity domain texts tend to receive slightly lower average scores compared to general-purpose educational texts. Therefore, we adopt a slightly more permissive threshold of 2.2. Combining these two sources yields a total of 500k seed documents.
Medical For the medical domain, we utilize the Biomed-Enriched dataset (Touchent et al., 2026), which annotates the PubMed Open Access (PMC-OA) corpus with document types, domains, and educational quality scores. We use the commercial split of this dataset and filter the seed documents based on the following criteria: an educational score of 4 or higher, domains categorized as either clinical or biomedical, and document types restricted to reviews and clinical cases. Furthermore, we exclude the "reference" sections and require all texts to have a minimum length of 200 characters. This filtering process results in a total of 450k seed documents.
Details of FIM-MCQA Synthesis
Our data construction pipeline converts raw documents into verifiable training tasks through a multi-stage process.
In the first phase, we utilize the gpt-oss-120b (OpenAI et al., [31]), configured with a medium reasoning effort, to simultaneously extract an educational passage and identify the target span for masking. If the document lacks educational content, the model is instructed to generate an empty response, ensuring this data is filtered out and does not proceed to the next phase.
Next, a deterministic programmatic verifier processes the passage and the identified span to construct the masked text. This programmatic layer applies a 4-tier matching protocol, consisting of exact, strip, NFKC, and fuzzy matching, to locate the span within the passage. It then replaces the verified substring with a [MASK] marker to produce the masked passage . If the substring check fails across all formatting tiers, the instance is rejected.
Finally, the gpt-oss-120b (OpenAI et al. [31]) (again with medium reasoning effort) generates the multiple-choice distractors. The model is prompted to create distractors such that each question contains a total of nine options, including the correct answer.
To ensure the quality of the constructed dataset, we apply a difficulty filtering step using the Qwen3-4b-Instruct-2507 (Yang et al. (2025)). Specifically, for each question, we sample the model’s responses 8 times, setting the temperature to 0.7 and the maximum token length to 6000. The statistics of the final filtered dataset are detailed in Appendix § 9.§ 4.
For the exact instructions, please refer to the detailed prompt templates provided in Figure 8, Figure 9, Figure 10, Figure 11.

Figure 8. The prompt template used in Stage 1 to simultaneously extract educational passages and identify mask spans for the Cybersecurity domain.

Figure 9. The prompt template used in Stage 1 to simultaneously extract educational passages and identify mask spans for the Medical domain.

Figure 10. The prompt template used in Stage 3 to generate plausible, length-matched distractors for the masked spans in the Cybersecurity domain.

Figure 11. The prompt template used in Stage 3 to generate plausible, length-matched distractors for the masked spans in the Medical domain.
Details of Cross-lingual Task Variant Synthesis
Details of Translation
We synthesize the TGT-ONLY data by fully translating the EN-ONLY dataset. For the translation process, we utilize the Gemma-4-26B-A4B-it (Gemma Team and Google DeepMind [13]) model configured with thinking=off. We translate the passages and multiple-choice options separately. This separate processing ensures the strict preservation of the [MASK] tokens and addresses the empirical tendency of large language models to under-translate or truncate long contexts. To mitigate this generation issue, passages are segmented and translated on a sentence-by-sentence basis, and the options are similarly split and translated individually. Furthermore, we provide few-shot exemplars within the prompt to enhance overall translation quality. Domain-specific terminology that must remain untranslated is managed separately via literal passthrough rules. The exact translation prompts are provided in Figure 12 and Figure 13.

Figure 12. The prompt template used for context-aware translation into the target language. The prompt enforces strict constraints to preserve code snippets, system identifiers, and [MASK] markers while preventing unnatural mid-word language mixing.

Figure 13. The prompt template used to translate the multiple-choice options into the target language. Similar to the passage translation, it leverages the source document as context to ensure terminology consistency and requires a strict JSON array output.
Details of Code-switching Process
We construct the option-side code-switching data by cross-matching the passages and multiple-choice options from the TGT-ONLY and EN-ONLY datasets. For the passage-level code-switched data,
we utilize the Gemma-4-26B-A4B-it (Gemma Team and Google DeepMind [13]) model configured with thinking=off. The code-switching pipeline operates by prompting the model to replace selected segments of a source passage with their target language counterparts. The model is strictly constrained to ensure contextual fluency, uniform span dispersion, and the exact preservation of [MASK] markers. The full prompt template is described in Figure 16.

Figure 16. The prompt template used for generating span-level code-switched passages.
Following the generation, we enforce a post-hoc filtering step: if the cumulative character length of the switched spans constitutes less than 15% of the total passage length, the generation is retried up to three times, with the decoding temperature increased by 0.3 for each sequential attempt.
Final data statistics
Table 6 summarizes the synthesized data pool by domain. We refer to this as a synthesized pool rather than the final training set, because our training runs use controlled subsamples and mix them with auxiliary data as described in § 4.
| Domain | Source | |
| Cybersecurity | 207,048 | Primus-Seed, Primus-FineWeb |
| Medical | 158,714 | Biomed-Enriched |
| Total | 365,762 |
Table 6. Synthesized FIM-MCQA data pool per domain. Note that for each domain-target language pair, an identical number of samples is constructed across all training variants, matching the exact counts reported in the table. The subsample used for training is described in Appendix § 10.§ 1.
Qualitative Examples
In this section, we present qualitative examples from our dataset. The examples for the medical domain are detailed in Figure 4, Figure 5, and those for the cybersecurity domain are provided in Figure 6, and Figure 7.

Figure 4. Qualitative example — Medical (Part 1/2). Same source passage shown across 3 variants. [MASK] marks the masked answer span; the ground-truth option (✓) is what fills it.

Figure 5. Qualitative example — Medical (Part 2/2). Continuation of the 3 variants.

Figure 6. Qualitative example — Cybersecurity (Part 1/2). Same source passage shown across 4 variants. [MASK] marks the masked answer span; the ground-truth option (✓) is what fills it.

Figure 7. Qualitative example — Cybersecurity (Part 2/2). Continuation of the 4 variants.
Experimental Setup
Additional Training Details
Model-specific effective sample filtering. To isolate effective training instances without compromising response lengths, we apply a model-specific filtering procedure to the synthesized FIM-MCQA data pool for each domain–target-language pair. Specifically, for all training variants, we evaluate the respective models on the EN-ONLY data by sampling responses eight times per instance. We discard instances where the model answers correctly across all eight trials, as well as those with response lengths falling into the bottom 25th percentile. After applying this filtering process, we retain only the intersection of the surviving instances across all models to construct the final base dataset. The filtered data counts per model and domain are detailed in Table 7.
| Domain | Model | Pool | Filtered |
| Medical | Qwen2.5-3B-Instruct | 158,714 | 50,476 |
| Medical | Llama-3.2-3B-Instruct | 158,714 | 44,294 |
| Cybersecurity | SmolLM3-3B | 207,048 | 43,705 |
Table 7. Number of Effective Samples after the difficulty filtering stage.
Training data mixture. Each single-view variant utilizes the entirety of this filtered data for its corresponding language configuration. For multi-view variants, we sample uniformly across the constituent views so that the total contribution of each view is balanced. Note that the same FIM-MCQA instance preserves its answer index across its monolingual, bilingual, and code-switched realizations, enabling all variants to share the identical exact-match verifier.
Auxiliary General-Reasoning dataset. The auxiliary RLVR pool is shared across all RL-trained variants and consists of 25K scientific reasoning problems [1], 15K math problems [27], and 10K instruction-following prompts [5]. For domain-RLVR variants, each batch samples prompts from the synthesized domain FIM-MCQA pool with probability 0.6 and from the auxiliary pool with probability 0.4. The RLVR-BASE baseline is trained only on the auxiliary pool. We use the same auxiliary pool and sampling rule across all domain variants so that differences among EN-ONLY, TGT-ONLY, BI-MONO, and OURS reflect the language realization of the domain examples rather than changes in generic RLVR data.
DAPO hyperparameters. We post-train all RL variants with DAPO (Yu et al., 2025a). For each prompt, we sample rollouts and apply accuracy-based group filtering, keeping a group only when it contains at least one correct and one incorrect answer. We disable the KL penalty and the overlong-buffer component. Rollouts are sampled with temperature 0.6 and top-, using a maximum prompt length of 4,200 tokens and a maximum response length of 8,192 tokens. We train for two epoch with a batch size of 256 prompts and a PPO mini-batch size of 64. Each variant requires approximately 2,500 H100 GPU hours.
Validation and Evaluation. For cybersecurity, we reserve 10% of CyberMetric for validation. For medicine, we reserve an exclusive 10% split of translated MedQA for validation. These validation sets are used only for hyperparameter selection and are excluded from final evaluation. The primary metric is pass@1 with , temperature , and top-.
Details of Evaluation Dataset Translation
We utilize the GPT-5.4-mini [33]5 model to perform the translation. The prompt template for translating the evaluation data is provided in Figure 14.

Figure 14. The prompt template used to translate evaluation benchmark multiple-choice questions and options into the target language.
Translation Quality Assessment of the Evaluation Dataset
To ensure the reliability of our cross-lingual evaluation, we assess the translation quality of the three cybersecurity benchmarks: SecEval [23], CyberMetric (Tihanyi et al., 2024), and CTI-MCQ [2]. The evaluation process incorporates an LLM-as-a-judge approach, qualitative human assessment, and quantitative automated metrics.
To evaluate the translations comprehensively, we employ GPT-5.5 as an LLM judge to score the translated texts on a scale of 1 to 5 across three criteria: Accuracy, Fluency, and Terminology. Drawing inspiration from Feng et al. [12], we specifically adjusted our prompt to elicit explicit numerical scores (on a scale of 1 to 5) in a strict JSON format. This design choice enables a direct quantitative comparison to calculate the agreement between the LLM and human judgments. To validate the reliability of this LLM-based evaluation, we measure its alignment with human judgments. We randomly
subsampled 20 instances from each benchmark, and three independent evaluators with Computer Science backgrounds assessed them. To ensure a fair and consistent comparison between the machine and human assessments, both GPT-5.5 (OpenAI, [33]) and the human experts were provided with the exact same instructions. The prompt utilized for this evaluation is illustrated in Figure 19. It asks the evaluator to output a strict JSON format scoring the translation on Accuracy, Fluency, and

Evaluation Data Translation Quality Assessment Prompt
Evaluate the translation on three aspects (1=worst, 5=best):
1. Accuracy: Does the translation preserve the source meaning?
2. Fluency: Is the translation grammatically correct and natural?
3. Terminology: Are domain-specific terms translated correctly?
Domain: {domain}
Source: {src}
Translation: {tgt}
Output JSON only:
{"accuracy": <1-5>, "fluency": <1-5>, "terminology": <1-5>}Figure 19. The prompt template used for both the LLM-as-a-judge (GPT-5.5) and the human evaluators to assess translation quality.
Terminology.
As shown in Table 8, GPT-5.5 yielded high scores across all benchmarks. When compared to the human evaluation, we observed a strong positive correlation, with Pearson values ranging from 0.79 to 0.81 across the datasets. This significant alignment demonstrates that our LLM-based evaluation pipeline reliably mirrors human judgment, effectively verifying the translation quality.
| Benchmark | GPT-5.5 | Human | KO | SW | |
| SecEval | 4.94/4.82/4.70 | 4.25/4.18/4.12 | 0.81 | 0.852 | 0.803 |
| CyberMetric | 4.92/4.75/4.67 | 4.21/4.14/4.08 | 0.83 | 0.864 | 0.810 |
| CTI-MCQ | 4.88/4.74/4.66 | 4.16/4.10/4.02 | 0.79 | 0.851 | 0.808 |
Table 8. Translation quality evaluation across cybersecurity benchmarks. GPT-5.5 and human scores represent assessments on a 20-instance subsample for Korean (evaluated on a 1–5 scale). denotes the Pearson correlation between LLM and human judgments. COMET scores are reported for the full evaluation sets in Korean and Swahili.
In addition to the prompt-based evaluation, we perform a comprehensive automated evaluation using the COMET score ((Rei et al., 2020)) to measure the semantic alignment between the original English texts and their translated counterparts. Specifically, we utilize the Unbabel/wmt22-cometkiwi-da model and compute the scores for the full evaluation sets across both target languages, Korean and Swahili, to ensure broad quantitative validation.
The detailed results of the LLM evaluation, human assessment, and the COMET automated metric are summarized in Table 8.
Additional Evaluation Results
Full Evaluation Results
In this section, we report the full evaluation results, including the no-RLVR settings. The results for medical and cybersecurity domains are presented in Table 9 and Table 10.
| Bench. | Qwen2.5-3B-Instruct | Llama-3.2-3B-Instruct | ||||||||||
| Base | RLVR-Base | +En | +Tgt | +BiMono | +Ours | Base | RLVR-Base | +En | +Tgt | +BiMono | +Ours | |
| Target Language | ||||||||||||
| KorMedMQ-D | 34.50 | 35.06 | 35.34 | 35.29 | 35.34 | 36.67 | 19.02 | 21.10 | 27.93 | 27.47 | 28.79 | 29.31 |
| MedMCQA* | 33.24 | 34.52 | 33.96 | 34.03 | 33.72 | 34.86 | 19.85 | 24.26 | 27.12 | 27.36 | 29.65 | 29.22 |
| KMMLU-Pro-M | 35.20 | 35.45 | 35.76 | 36.21 | 36.06 | 36.29 | 20.45 | 22.84 | 24.77 | 24.09 | 25.08 | 26.36 |
| Avg (Tgt) | 34.31 | 35.01 | 35.02 | 35.18 | 35.04 | 35.94 | 19.77 | 22.73 | 26.61 | 26.31 | 27.84 | 28.30 |
| English | ||||||||||||
| KorMedMQ-D* | 57.53 | 57.74 | 58.05 | 61.03 | 60.06 | 58.74 | 48.85 | 50.01 | 56.72 | 57.30 | 57.53 | 57.93 |
| MedMCQA | 47.36 | 47.88 | 47.60 | 47.42 | 47.34 | 47.01 | 52.28 | 52.57 | 55.00 | 55.00 | 55.17 | 55.04 |
| MedQA | 50.29 | 48.65 | 50.92 | 50.47 | 51.47 | 50.65 | 50.63 | 52.38 | 58.19 | 58.37 | 58.92 | 58.82 |
| Avg (En) | 51.73 | 51.42 | 52.19 | 52.97 | 52.96 | 52.13 | 50.59 | 51.65 | 56.64 | 56.89 | 57.21 | 57.26 |
Table 9. Comprehensive performance on medical benchmarks across RLVR data configurations. BASE denotes the original instruction-tuned model without RLVR. RLVR-BASE denotes RLVR training using only the auxiliary data pool. The remaining variants include the same auxiliary pool and additionally use English data (+EN), target-language data (+TGT), bilingual monolingual data (+BIMONO), or our proposed code-switched data (+OURS). English rows are shown in gray to indicate source-language evaluation, which is reported as a retention check rather than the primary target-language transfer metric. The asterisk (*) indicates that the benchmark is translated into the respective language. The best performance within each model is highlighted in bold.
| Bench. | SmolLM3-3B Korean | SmolLM3-3B Swahili | ||||||||||
| Base | RLVR-Base | +En | +Tgt | +BiMono | +Ours | Base | RLVR-Base | +En | +Tgt | +BiMono | +Ours | |
| Target Language | ||||||||||||
| SecEval* | 40.9 | 44.3 | 50.1 | 49.8 | 51.1 | 57.3 | 26.9 | 25.8 | 26.8 | 28.6 | 25.9 | 28.3 |
| CyberMetric* | 63.6 | 65.8 | 70.6 | 80.2 | 72.4 | 80.0 | 53.0 | 54.0 | 53.9 | 54.2 | 54.5 | 55.0 |
| CTI-MCQ* | 33.7 | 40.6 | 43.2 | 45.5 | 43.8 | 47.1 | 41.4 | 41.9 | 42.1 | 43.2 | 43.4 | 45.0 |
| Avg (Tgt) | 46.07 | 50.23 | 54.63 | 58.50 | 55.77 | 61.47 | 40.43 | 40.57 | 40.93 | 42.00 | 41.27 | 42.77 |
| English | ||||||||||||
| SecEval | 55.6 | 56.7 | 59.9 | 61.7 | 58.9 | 62.7 | 55.6 | 56.7 | 59.9 | 65.6 | 65.3 | 66.6 |
| CyberMetric | 77.0 | 83.0 | 83.6 | 85.6 | 84.4 | 85.6 | 77.0 | 83.0 | 83.6 | 86.3 | 86.0 | 86.6 |
| CTI-MCQ | 50.7 | 54.9 | 57.5 | 56.2 | 57.1 | 56.5 | 50.7 | 54.9 | 57.5 | 61.1 | 60.9 | 60.3 |
| Avg (En) | 61.10 | 64.87 | 67.00 | 67.83 | 66.80 | 68.27 | 61.10 | 64.87 | 67.00 | 71.00 | 70.73 | 71.17 |
Table 10. Comprehensive performance on cybersecurity benchmarks across RLVR data configurations. Target Language indicates Korean for the left section and Swahili for the right section. BASE denotes the original instruction-tuned model without RLVR. RLVR-BASE denotes RLVR training using only the auxiliary data pool. The remaining variants include the same auxiliary pool and additionally use English data (+EN), target-language data (+TGT), bilingual monolingual data (+BIMONO), or our proposed code-switched data (+OURS). English rows are shown in gray to indicate source-language evaluation, which is reported as a retention check rather than the primary target-language transfer metric. The asterisk (*) indicates that the benchmark is translated into the respective language. The best performance within each model is highlighted in bold.
Multi-Seed Robustness
We extend the Korean cybersecurity comparison with SmolLM3-3B to three random seeds. Table 11 reports the mean and standard deviation for the five RL-trained variants. The same performance pattern is observed across the three seeds. +OURS achieves the highest target-language average, performs best on SecEval and CTI-MCQ, and matches +TGT on CyberMetric.
| Variant | SecEval | CyberMetric | CTI-MCQ | |
| RLVR-BASE | ||||
| +EN | ||||
| +TGT | ||||
| +BIMONO | ||||
| +OURS |
Table 11. Multi-seed results for SmolLM3-3B on Korean cybersecurity benchmarks, evaluated in the target language. Values report mean standard deviation over three random seeds. Best results are bolded.
Detailed Component Ablation Results
In this section, we present the component ablation study results across individual benchmarks. The complete results are detailed in Table 12.
| Bench. | RLVR-Base | +En | +Tgt | +BiMono | +Opt | +Pass | +Ours |
| Target Language | |||||||
| SecEval* | 44.3 | 50.1 | 49.8 | 51.1 | 52.0 | 54.7 | 57.3 |
| CyberMetric* | 65.8 | 70.6 | 80.2 | 72.4 | 71.0 | 77.4 | 80.0 |
| CTI-MCQ* | 40.6 | 43.2 | 45.5 | 43.8 | 45.9 | 46.0 | 47.1 |
| Avg (Tgt) | 50.23 | 54.63 | 58.50 | 55.77 | 56.30 | 59.37 | 61.47 |
| English | |||||||
| SecEval | 56.7 | 59.9 | 61.7 | 58.9 | 60.3 | 61.2 | 62.7 |
| CyberMetric | 83.0 | 83.6 | 85.6 | 84.4 | 84.0 | 84.0 | 85.6 |
| CTI-MCQ | 54.9 | 57.5 | 56.2 | 57.1 | 57.3 | 56.0 | 56.5 |
| Avg (En) | 64.87 | 67.00 | 67.83 | 66.80 | 67.20 | 67.07 | 68.27 |
Table 12. Component ablation on Korean cybersecurity benchmarks with SmolLM3-3B. OPT-ONLY and PASS-ONLY extend BI-MONO with option-side and passage-level code-switching, respectively, while OURS combines both components. All variants follow the same experimental setup. Asterisks denote target-language versions of the benchmarks.
Analysis Details
Details of Cross-lingual Consistency Comparison
In this section, we present a detailed breakdown of the results from Table 4 across individual benchmarks. These results are provided in Table 13.
| (English, Korean) | (✓,✓) ↑ | (✓,×) ↓ | (×,✓) ↓ |
| SecEval | |||
| Base | 8.3 | 39.2 | 5.4 |
| En-only | 15.3 | 41.8 | 3.4 |
| Bi-Mono | 16.3 | 41.1 | 3.7 |
| Ours | 23.8 | 38.8 | 2.5 |
| CyberMetric | |||
| Base | 31.4 | 13.8 | 32.0 |
| En-only | 62.0 | 15.6 | 9.6 |
| Bi-Mono | 64.0 | 17.6 | 7.2 |
| Ours | 74.4 | 11.2 | 5.4 |
| CTI-MCQ | |||
| Base | 17.7 | 21.4 | 15.8 |
| En-only | 32.6 | 22.6 | 10.4 |
| Bi-Mono | 33.5 | 22.4 | 10.0 |
| Ours | 36.5 | 20.0 | 10.6 |
Table 13. Detailed cross-lingual consistency comparison for each benchmark. Bold indicates the highest consistency.
Details of In-domain Translation Experiment
The performance ratio of variant on benchmark is , where is the variant’s accuracy on the items of rendered in language , evaluated on the same stem-key paired item set. The translation prompts used by each model are provided in Figure 15. Figure 17 shows the exact prompt template used for the LLM-based direct assessment (GEMBA-DA) with GPT-4.1 ([32]) to evaluate translation quality from the target language to English.

Figure 15. The prompt template used for translating source language text into English.

GEMBA-DA Evaluation Prompt
Score the following translation from {src_full} to {tgt_full} on a
continuous scale from 0 to 100, where a score of zero means “no meaning
preserved” and score of one hundred means “perfect meaning and grammar”.
{src_full} source: “{src}”
{tgt_full} translation: “{tgt}”
Score (0-100):Figure 17. The prompt template for GEMBA-DA used to measure translation quality. The variables {src_full} and {tgt_full} represent the full names of the source and target languages, while {src} and {tgt} are substituted with the original text and the model’s translation, respectively.
Details of Representation Alignment Analysis
We use a fixed subset of semantically matched source–target question pairs from CyberMetric. The source side is the original English question, and the target side is the Korean translation used in our evaluation. For each item, we render the same multiple-choice prompt in both languages and append the same opening reasoning marker, <think>.
We run a single forward pass over the prompt plus the opening reasoning marker and extract the hidden state at the marker position from every layer. If the marker is split into multiple tokenizer tokens, we use the hidden state of the final token of the marker. This position is immediately before the first reasoning token would be predicted. Since the marker is identical across the source and target prompts, differences in its contextualized hidden state mainly reflect how the preceding question is represented before reasoning begins, rather than variation from generated chain-of-thought continuations.
For each item and layer , let and denote the hidden states extracted from the source and target prompts. Stacking them over all paired items gives layer-wise matrices . We mean-center each language separately over the item dimension:
We compute two alignment metrics. First, Mean-Centered Cosine measures paired item-level similarity after removing the per-language mean, which reduces the effect of anisotropy in contextual LM representations [9] (Timkey and van Schijndel 2021):
Second, ANC measures feature-level activation correspondence by averaging the Pearson correlation of each hidden dimension over the aligned items [7]:
We report the full layer-wise curves and summarize the results over a fixed middle-layer band, layers 12–24 of the 36-layer model. The main text focuses on this band because intermediate representations are commonly used to study cross-lingual alignment, and because this range captures the largest separation between the code-switched model and the non-code-switched baselines in our analysis.
Details of Generated Reasoning Trace Alignment
For each benchmark item, we construct an English–target-language pair and collect outputs from each evaluated model under both language conditions. We compare Ours against En-only and Tgt-only using GPT-5.5 as a blind pairwise judge with reasoning effort set to medium. The judge receives anonymized outputs from two systems and selects the system whose English and target-language outputs are more semantically consistent in high-level logical progression, problem decomposition, intermediate conclusions, and solution strategy. The judge is instructed to ignore surface-level differences such as wording, formatting, verbosity, fluency, and language-specific phrasing.
To reduce position bias, we evaluate each comparison twice with the A/B order reversed. We aggregate the two judgments into a single outcome per example. If both orders prefer the same underlying system, we count it as a win for that system. We count an example as a tie only when both judgments explicitly return a tie. Unless otherwise specified, denotes the number of examples before A/B order swapping.
For the medical domain, we evaluate Qwen2.5-3B-Instruct and Llama-3.2-3B-Instruct on four benchmarks with 100 randomly sampled examples per benchmark, resulting in examples per model family. For cybersecurity, we evaluate SmolLM3-3B on three benchmarks with 100 examples per benchmark, resulting in . The judge prompt is shown in Figure 18, and the per-benchmark results are reported in Table 14.
| Benchmark | Ours% | En-only% | tie% | Ours% | Tgt-only% | tie% | |
| Medical | |||||||
| Qwen2.5-3B | MedQA | 47.0 | 36.0 | 1.0 | 48.0 | 41.0 | 0.0 |
| Qwen2.5-3B | MedMCQA | 50.0 | 40.0 | 3.0 | 46.0 | 47.0 | 1.0 |
| Qwen2.5-3B | KMMLU-Pro-Medical | 50.0 | 37.0 | 0.0 | 54.0 | 36.0 | 0.0 |
| Qwen2.5-3B | KorMedMCQA-Doctor | 46.0 | 47.0 | 0.0 | 42.0 | 46.0 | 2.0 |
| Qwen2.5-3B | Average | 48.3 | 40.0 | 1.0 | 47.5 | 42.5 | 0.8 |
| Llama-3.2-3B | MedQA | 44.0 | 45.0 | 3.0 | 55.0 | 39.0 | 0.0 |
| Llama-3.2-3B | MedMCQA | 41.0 | 33.0 | 5.0 | 30.0 | 48.0 | 3.0 |
| Llama-3.2-3B | KMMLU-Pro-Medical | 48.0 | 43.0 | 1.0 | 46.0 | 44.0 | 0.0 |
| Llama-3.2-3B | KorMedMCQA-Doctor | 55.0 | 34.0 | 0.0 | 42.0 | 37.0 | 0.0 |
| Llama-3.2-3B | Average | 47.0 | 38.8 | 2.3 | 43.3 | 42.0 | 0.8 |
| Cyber | |||||||
| SmolLM3-3B | CTI-MCQ | 45.0 | 35.0 | 5.0 | 69.0 | 26.0 | 0.0 |
| SmolLM3-3B | CyberMetric | 31.0 | 30.0 | 13.0 | 25.0 | 27.0 | 17.0 |
| SmolLM3-3B | SecEval | 33.0 | 30.0 | 12.0 | 82.0 | 12.0 | 0.0 |
| SmolLM3-3B | Average | 36.3 | 31.7 | 10.0 | 58.7 | 21.7 | 5.7 |
Table 14. Detailed pairwise judge win rate results using GPT-5.5. The target language is Korean. 100 samples are evaluated per benchmark. Bold indicates a higher win rate.

Reasoning Trace Alignment Judge Prompt
You will see TWO models' reasoning behavior on the same problem.
Each model produced one English trace (for the English version
of the problem) and one {target_name} trace (for the
{target_name} version of the same problem).
Your task: judge which MODEL shows MORE CONSISTENT REASONING
between its English and {target_name} responses.
If the two models show a similar degree of reasoning consistency, choose TIE. Do not force a
choice between the two models when there is no meaningful difference in consistency.
WHAT "CONSISTENT" MEANS:
- The model decomposes the problem the same way in both languages
- Uses similar key facts, concepts, formulas across both traces
- Follows a similar high-level strategy in both languages
- Arrives at similar intermediate insights
WHAT "CONSISTENT" DOES NOT MEAN:
- Both traces being correct (a model can be consistently wrong,
and that still counts as consistent)
- The {target_name} trace using English words (this is a surface
feature, not consistency)
- Either trace being more detailed or longer than the other
- Either model being individually "smarter" or producing
higher-quality reasoning
PROBLEM (English):
{Q_en}
PROBLEM ({target_name}):
{Q_tgt}Figure 18. The prompt template used for the reasoning trace alignment experiment.
Prompts
In this section, we present the full prompt templates used in our experiments. To facilitate navigation, we group these prompts according to their specific tasks: extraction synthesis, translation, code-switching, and evaluation.
Extraction Synthesis Prompts
The prompts for extracting educational passages and generating distractors across domains are provided in Figure 8, Figure 9, Figure 10, and Figure 11.
Translation Prompts
The prompts utilized for context-aware passage translation, option translation, evaluation benchmark translation, and direct translation are shown in Figure 12, Figure 13, Figure 14, and Figure 15.
Code-Switching Prompt
The prompt template used for span-level code-switching generation is detailed in Figure 16.
Evaluation and Judging Prompts
The prompts used for evaluation tasks, including GEMBA-DA, reasoning trace alignment, and translation quality assessment, are presented in Figure 17, Figure 18, and Figure 19.
Usage of AI assistants
During the preparation of this manuscript, AI-based writing tools were used solely to improve the clarity, grammar, and readability of the English text. They were not involved in formulating the research questions, developing the methodology, conducting the experiments, or interpreting the results. All technical content, analyses, and conclusions were independently developed by the authors.
Variant: En-only
Passage: Anti-programmed death-1 (anti-PD-1) and programmed death-ligand 1 (anti-PD-L1) monoclonal antibodies are approved treatments for several malignancies, including nivolumab and pembrolizumab for melanoma and non-small-cell lung cancer (NSCLC), nivolumab for renal cell carcinoma and Hodgkin’s lymphoma, atezolizumab for bladder cancer, and the combination of nivolumab plus ipilimumab for melanoma. While these agents are generally associated with mild toxicity, they can provoke serious immune-related adverse events. [MASK]. Drug-induced pneumonia may also arise from conventional chemotherapy agents such as docetaxel, gemcitabine, and bleomycin, from targeted therapies like epidermal growth factor receptor inhibitors and mammalian target of rapamycin inhibitors, as well as from radiotherapy.
Options:
A) Pneumonia—defined as focal or diffuse inflammation of the lung parenchyma—is one such adverse event; its reported incidence in clinical trials of anti-PD-1/PD-L1 antibodies ranges from 0 % to 10 % ✓
B) Colitis — characterized by ulcerative lesions of the intestinal mucosa — another potential toxicity ; its observed frequency in phase III studies of PD-1/PD-L1 blockers varies between 2 % and 12 % in treated patients
C) Pancreatitis involving inflammatory swelling of the exocrine gland is a rare metabolic side effect ; its noted proportion within large anti-PD-1/PD-L treatment groups hovers around three percent through eight percent across cycles
D) Hypothyroidism — characterized by reduced thyroid hormone production and fatigue — a common endocrine immune-related adverse effect ; its occurrence rate across anti-PD-1/PD-L1 studies lies near 15 % to 30 % globally in adult populations
E) Myocarditis — manifested as inflammatory infiltration of cardiac muscle — an infrequent yet severe checkpoint inhibitor toxicity ; its prevalence reported in extensive PD-1/PD-L1 cohorts fluctuates between 1 % and 5 % overall in patients
F) Hepatitis — identified by elevated transaminases and hepatic inflammation — a documented immune-related complication ; its cumulative incidence across pivotal PD-1/PD-L1 trials spans roughly from 3 % up to 9 % among study participants
G) Dermatitis — described as erythematous rash covering variable body surface area — a frequent cutaneous manifestation ; its incidence reported in major anti-PD-01/PD-L1 investigations spans roughly from 20 % to 45 %, depending on regimen
H) Nephritis — presenting with proteinuria haematuria and impaired renal function — another recognized renal toxicity ; its documented frequency among PD-1/PD-L1 recipients ranges approximately from 2 % up to 8 % during therapy in clinical trials
I) Encephalitis — defined as brain parenchymal inflammation with neurological symptoms — an uncommon neurotoxic event ; its estimated rate among pooled PD-1/PD-L1 patient data sets runs from five per cent up to fifteen per cent during follow up
Usage of AI assistants
Extraction Synthesis Stage 1 Prompt (Cybersecurity)
Given a source document, perform TWO tasks:
1. **Extract an educational passage**: Extract a coherent, self-contained passage (around 200–300 words) from the document that is educationally valuable for learning Cybersecurity. The passage should contain reasoning - causal explanations, attack mechanism analysis, vulnerability consequences, or defensive logic. Remove noise such as ads, navigation menus, author bylines, dates, and promotional content. The passage must read as a clean, standalone educational text.
2. **Identify the mask span**: Within the extracted passage, identify a consecutive multi-sentence span (around 50–100 words) that contains the most important reasoning steps. This span will be masked and serve as the ground-truth answer.
VERBATIM COPYING RULES – violating these causes a task failure:
1. Both “passage” and “removed_steps” must be copied CHARACTER-FOR-CHARACTER from the original document.
2. Do NOT paraphrase, simplify, reorder, or summarize.
3. Do NOT add or remove spaces, line breaks, or punctuation.
4. “removed_steps” must be a substring of “passage” – every character must appear in the passage exactly as written.
Return ONLY the following JSON within <answer></answer> tags:
<answer>
{
"passage": "...",
"removed_steps": "..."
}
</answer>
If there is no suitable educational passage about Cybersecurity, or if the document is not written in English, return an empty string within <answer></answer>.
Document:
SOLUTION_PLACEHOLDER
Direct Translation Prompt
Translate the following {src_lang_full} text to English.
Output only the English translation, no explanation, no preamble.
{src_lang_full} text:
{source_text}
Span-Level Code-Switching Generation Prompt
You will code-switch a {source} ({source_native}) {domain} text into {target}
({target_native}) by replacing selected spans with their {target} translations. Choose
spans so that the resulting text READS NATURALLY - the output should feel like a fluent
bilingual {domain} writer's prose, not a mechanical word-by-word substitution.
Rules (STRICT):
1. Pick spans whose {target} translation READS NATURALLY in context. Translate
content-carrying spans - nouns, verbs, adjectives, short noun/verb phrases - whenever
their {target} rendering is smooth. Spans may be single words or short 2-4 token phrases -
phrasal translations often read more naturally than single-word swaps, so prefer them when
they flow better.
2. Domain literals stay in their original form (do NOT translate or transliterate):
{literal_passthroughs}, code blocks, equations.
3. Spread selections EVENLY across the whole passage - the beginning, the middle, and the end
must each contribute some switched spans. Clustering in one region is a failure.
4. If the same word appears several times, you do NOT have to translate every occurrence.
Translating one or two occurrences and leaving the others in {source} is actually preferred
- it looks more natural and avoids the "always the same word" bias.
5. Translate each selected span into natural {target} and substitute it in place of the
{source} original. Keep the surrounding {source} text exactly as it was (same whitespace,
same casing, same punctuation, same paragraph breaks).
6. NEVER modify, translate, remove, or reorder the markers [MASK]. They must appear the same
number of times and in the same order in your output.
7. Do not add, remove, or reorder sentences. Do not paraphrase untouched text.
{example_block}
OUTPUT FORMAT - respond with ONLY the following, nothing else:
<answer>
{
"code_switched": "<the full passage after code-switching>",
"switched_tokens": ["<{source} span 1>", "<{source} span 2>", ...]
}
</answer>
Emit exactly ONE opening { and ONE matching closing }. Do NOT double them.
Escape newlines inside the JSON string as \n and backslashes as \\.
PASSAGE:
<<<
PASSAGE_PLACEHOLDER
>>>References
- [1]Syeda Nahida Akter, Shrimai Prabhumoye, Matvei Novikov, Seungju Han, Ying Lin, Evelina Bakhturina, Eric Nyberg, Yejin Choi, Mostofa Patwary, Mohammad Shoeybi, and Bryan Catanzaro. 2026. Nemetron-CrossThink: Scaling self-learning beyond math reasoning. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 984–1002, Rabat, Morocco. Association for Computational Linguistics.DOI
- [2]Md Tanvirul Alam, Dipkamal Bhusal, Le Nguyen, and Nidhi Rastogi. 2024. Citbench: A benchmark for evaluating llms in cyber threat intelligence. In Advances in Neural Information Processing Systems, volume 37, pages 50805–50825. Curran Associates, Inc.
- [3]Peter Auer. 1998. Code-Switching in Conversation: Language, Interaction and Identity. Routledge, London.
- [4]Elie Bakouch, Loubna Ben Allal, Anton Lozhkov, Nouamane Tazi, Lewis Tunstall, Carlos Miguel Patiño, Edward Beeching, Aymeric Roucher, Aksel Joonas Reedi, Quentin Gallouédec, Kashif Rasul, Nathan Habib, Clémentine Fourrier, Hynek Kydlicek, Guilherme Penedo, Hugo Larcher, Mathieu Morlon, Vaibhav Srivastav, Joshua Lochner, and 4 others. 2025. SmolLM3: smol, multilingual, long-context reasoner. https://huggingface.co/blog/smollm3.
- [5]Akhiad Bercovich, Itay Levy, Izik Golan, Mohammad Dabbah, Ran El-Yaniv, Omri Puny, Ido Galil, Zach Moshe, Tomer Ronen, Najeeb Nabwani, Ido Shafaf, Oren Tropp, Ehud Karpas, Ran Zilberstein, Jiaqi Zeng, Soumye Singhal, Alexander Bukhorn, Yian Zhang, Tugrul Konuk, and 117 others. 2025. Llama-nemetron: Efficient reasoning models. Preprint, arXiv:2505.00949.arxiv.org/abs/2505.00949
- [6]Nuo Chen, Zinan Zheng, Ning Wu, Ming Gong, Dongmei Zhang, and Jia Li. 2024. Breaking language barriers in multilingual mathematical reasoning: Insights and observations. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7001–7016, Miami, Florida, USA. Association for Computational Linguistics.arxiv.org/abs/2310.20246
- [7]Maksym Del and Mark Fishel. 2022. Cross-lingual similarity of multilingual representations revisited. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 185–195, Online only. Association for Computational Linguistics.arxiv.org/abs/2212.01924
- [8]Konstantin Dobler, Simon Lehnerer, Federico Scozzafava, Jonathan Janke, and Mohamed Ali. 2026. macereason-math: A dataset of high-quality multilingual math problems ready for rlvr. Preprint, arXiv:2603.10767.arxiv.org/abs/2603.10767
- [9]Kawin Ethayarajh. 2019. How contextual are contextualized word representations? Comparing the geometry of BERT, ELMo, and GPT-2 embeddings. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 55–65, Hong Kong, China. Association for Computational Linguistics.arxiv.org/abs/1909.00512
- [11]Zhaopeng Feng, Shaosheng Cao, Jiahan Ren, Jiayuan Su, Ruizhe Chen, Yan Zhang, Jian Wu, and Zuozhu Liu. 2025a. MT-r1-zero: Advancing LLM-based machine translation via rl-zero-like reinforcement learning. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 18685–18702, Suzhou, China. Association for Computational Linguistics.arxiv.org/abs/2504.10160
- [12]Zhaopeng Feng, Jiayuan Su, Jiamei Zheng, Jiahan Ren, Yan Zhang, Jian Wu, Hongwei Wang, and Zuozhu Liu. 2025b. M-MAD: Multidimensional multi-agent debate for advanced machine translation evaluation. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7084–7107, Vienna, Austria. Association for Computational Linguistics.arxiv.org/abs/2412.20127
- [13]Gemma Team and Google DeepMind. 2026. Gemma 4 model card. https://ai.google.dev/gemma/docs/core/model_card_4. Accessed: 2026-05-26.
- [14]Aaron Grattafiori, Abhinav Pande, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravanakumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783.arxiv.org/abs/2407.21783
- [15]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Guo, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, and 175 others. 2025. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645(8081):633–638.DOI
- [16]Xu Guo, Qiming Ge, Jian Tong, Kedi Chen, Jin Zhang, Xiaogui Yang, Xuan Gao, Haijun Lv, Zhihui Lu, Yicheng Zou, and Qipeng Guo. 2026. Rethinking multiple-choice questions for RLVR: Unlocking potential via distractor design. In Findings of the Association for Computational Linguistics: ACL 2026, pages 20092–20113, San Diego, California, United States. Association for Computational Linguistics.
- [17]Seokhee Hong, Sunkyung Kim, Guijin Son, Soyeon Kim, Yeonjung Hong, and Jinsik Lee. 2025. From KMMLU-redux to pro: A professional Korean benchmark suite for LLM evaluation. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 19067–19096, Suzhou, China. Association for Computational Linguistics.
- [18]Jian Hu, Mingjie Liu, Ximing Lu, Fang Wu, Zaid Harchaoui, Shizhe Diao, Yejin Choi, Pavlo Molchanov, Jun Yang, Jan Kautz, and Yi Dong. 2026. BRoRL: Scaling reinforcement learning via broadened exploration. In Forty-third International Conference on Machine Learning.arxiv.org/abs/2510.01180
- [19]Di Jin, Eileen Pan, Nassim Oufattolle, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2020. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Preprint, arXiv:2009.13081.arxiv.org/abs/2009.13081
- [20]Deokhyung Kang, Seonjeong Hwang, DaeHui Kim, Hyounghun Kim, and Gary Lee. 2026. Why do multilingual reasoning gaps emerge in reasoning language models? In Findings of the Association for Computational Linguistics: ACL 2026, pages 31684–31716, San Diego, California, United States. Association for Computational Linguistics.
- [21]Hyunwoo Ko, Guijin Son, and Dasol Choi. 2025. Understand, solve and translate: Bridging the multilingual mathematical reasoning gap. In Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025), pages 78–95, Suzhou, China. Association for Computational Linguistics.arxiv.org/abs/2501.02448
- [22]Sunjun Kweon, Byungjin Choi, Gyouk Chu, Junyeong Song, Daeun Hyeon, Sujin Gan, Jueon Kim, Minkyu Kim, Rae Woong Park, and Edward Choi. 2024. Ko-medmcqa: Multi-choice question answering benchmark for korean healthcare professional licensing examinations. Preprint, arXiv:2403.01469.arxiv.org/abs/2403.01469
- [23]Guancheng Li, Yifeng Li, Wang Guannan, Haoyu Yang, and Yang Yu. 2023. SecEval: A comprehensive benchmark for evaluating cybersecurity knowledge of foundation models. https://github.com/XuanwuAI/SecEval.
- [24]Danni Liu and Jan Niehues. 2025. Middle-layer representation alignment for cross-lingual transfer in fine-tuned LLMs. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 15979–15996, Vienna, Austria. Association for Computational Linguistics.
- [25]Mingjie Liu, Shizhe Diao, Ximing Lu, Jian Hu, Xin Dong, Yejin Choi, Jan Kautz, and Yi Dong. 2025. Prol: Prolonged reinforcement learning expands reasoning boundaries in large language models. In Advances in Neural Information Processing Systems, volume 38, Main Conference, pages 17998–18031. Curran Associates, Inc.arxiv.org/abs/2505.24864
- [26]Ximing Lu, David Acuna, Jaehun Jung, Jian Hu, Di Zhang, Shizhe Diao, Yunheng Zou, Shaokun Zhang, Brandon Cui, Mingjie Liu, Hyunwoo Kim, Prithviraj Ammanabrolu, Jan Kautz, Yi Dong, and Yejin Choi. 2026. Golden goose: A simple trick to synthesize unlimited RLVR tasks from unverifiable internet text. In Forty-third International Conference on Machine Learning.
- [27]Michael Luo, Sijun Tan, Justin Wong, Xiaoxiang Shi, William Tang, Manan Roongta, Colin Cai, Jeffrey Luo, Tianjun Zhang, Erran Li, Racula Ada Popa, and Ion Stoica. 2025a. Deepscaler: Surpassing o1-preview with a 1.5b model by scaling rl. Notion Blog.
- [28]Wenyang Luo, Wayne Xin Zhao, Jing Sha, Shijin Wang, and Ji-Rong Wen. 2025b. MMATH: A multilingual benchmark for mathematical reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 11187–11202, Suzhou, China. Association for Computational Linguistics.arxiv.org/abs/2505.19126
- [29]Xueguang Ma, Qian Liu, Dongfu Jiang, Ge Zhang, Zejun MA, and Wenhu Chen. 2025. General-reasoner: Advancing llm reasoning across all domains. In Advances in Neural Information Processing Systems, volume 38, Main Conference, pages 56596–56618. Curran Associates, Inc.arxiv.org/abs/2505.14652
- [30]Benjamin Muller, Yanai Elazar, Benoît Sagot, and Djamé Seddah. 2021. First align, then predict: Understanding the cross-lingual ability of multilingual BERT. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 2214–2231. Online. Association for Computational Linguistics.arxiv.org/abs/2101.11109
- [31]OpenAI, ; Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwin Arbus, Rahul K. Arora, Yu Bai, Bowen Baker, Haiming Bao, Boaz Barak, Ally Bennett, Tyler Bertao, Nivedita Brett, Eugene Brevdo, Greg Brockman, Sebastien Bubeck, and 108 others. 2025. gpt-oss-120b & gpt-oss-20b model card. Preprint, arXiv:2508.10925.
- [32]OpenAI. 2025. Introducing GPT-4.1 in the API. https://openai.com/index/gpt-4-1/. Accessed: 2026-05-26.
- [33]OpenAI. 2026a. GPT-5.5 system card. Technical report, OpenAI. Accessed: 2026-05-26.

