Learning When to Translate for Multilingual Reasoning Deokhyung Kang, Hyounghun Kim, Gary Geunbae Lee Abstract Reasoning language models (RLMs) achieve strong performance on complex reasoning tasks, but still exhibit substantial multilingual reasoning gaps, largely due to language-understanding failures in non-English inputs. English translation can mitigate these failures by expressing non-English inputs in a form that RLMs can more reliably interpret, yet translating every input is unnecessary when the model can reason reliably from the original query. To address this challenge, we propose Luar , a L anguage U nderstanding Boundary- a ware R einforcement Learning framework that trains RLMs to selectively invoke translation when direct understanding is unreliable. Luar trains the model to choose between solving the original input directly and reasoning over its English translation, encouraging translation only when translator-augmented reasoning is expected to substantially outperform direct reasoning. Across multilingual reasoning benchmarks, Luar outperforms standard GRPO and other training-based baselines, with particularly large gains on low-resource languages. Further analysis shows that Luar avoids unnecessary translation in cases where direct reasoning is sufficient, while extending its translator-call behavior to unseen low-resource languages. Together, our work suggests a selective approach to multilingual reasoning: RLMs can learn to invoke translation only when their direct understanding is unreliable. 1 1 1 The project will be made publicly available at https://github.com/deokhk/LUAR . Figure 1: Explicit uncertainty rates of RLMs on PolyMath-Low. We report the percentage of query-understanding failures where the model explicitly expresses uncertainty about the meaning of the query, measured separately in reasoning traces and final responses. Evaluation details are provided in Appendix A . 1 Introduction Reasoning language models ( RLMs ) have made rapid progress on complex reasoning tasks by generating intermediate reasoning traces before producing final responses 20 ; 3 ; 32 . Despite this progress, RLMs still exhibit multilingual reasoning performance gaps 6 ; 30 . In particular, they perform much better on queries in high-resource languages such as English than in low-resource languages, making RLMs less reliable for speakers of underrepresented languages. Recent studies 11 ; 5 suggest that these gaps are largely driven by failures in language understanding, especially for low-resource languages. The key bottleneck is therefore often not reasoning itself, but accurately comprehending the meaning of the input. A straightforward approach is to directly improve the model’s language understanding across languages. However, this approach is difficult to scale across languages, especially for RLMs. Continual training on target-language data 37 ; 29 ; 13 requires targeted training data and optimization for each language. Model merging with multilingual models reduces language-specific data requirements 35 ; 7 ; 22 , but requires careful adaptation to post-trained RLMs to avoid reasoning performance degradation 33 . These limitations motivate our central research question: rather than directly improving the model’s own language understanding across languages, can RLMs learn to recognize when an input falls outside their own language understanding boundary and selectively invoke translation? To examine whether current RLMs exhibit signals relevant to this boundary, we analyze cases where incorrect answers stem from failures to understand the query. We measure the explicit uncertainty rate, the percentage of such cases where the model explicitly expresses uncertainty about its understanding. Figure 1 shows that models often reveal such uncertainty in reasoning traces, yet still produce confident final responses. This suggests that RLMs exhibit internal signals of language-understanding failure, but lack a mechanism to act on them. This motivates the idea of invoking English translation when needed, which can present the input in a more reliable form. To operationalize this idea, we propose Luar , a L anguage U nderstanding Boundary- a ware R einforcement Learning framework that teaches RLMs to invoke translation only when direct understanding is unreliable. Luar trains this decision inside the reasoning trace through two stages: translator-call supervised fine-tuning exposes the model to both direct and translator-augmented reasoning, and boundary-aware online reinforcement learning with GRPO 25 rewards translation only when it is expected to substantially improve final-answer accuracy. Across multilingual mathematical and STEM reasoning benchmarks 30 ; 31 , Luar improves Qwen3-4B and Qwen3-8B 32 over standard GRPO and other training-based baselines, with especially large gains on low-resource languages. Its learned policy concentrates translator calls on languages where translation is useful, while preserving direct reasoning for high-resource languages and generalizing to unseen low-resource languages such as Yoruba and Zulu. We summarize our main contributions as follows: (i) We introduce a view of multilingual reasoning centered on language-understanding boundaries, in which RLMs selectively invoke translation when direct understanding is unreliable. (ii) We propose Luar , a two-stage framework that learns when to invoke translation from outcome-based boundary signals. (iii) We demonstrate that LUAR substantially improves multilingual reasoning performance, especially for low-resource languages, by using translation selectively. 2 Related Work Despite the rapid progress of language models on complex reasoning tasks, they continue to exhibit multilingual reasoning gaps, especially for low-resource languages 6 ; 1 ; 12 . Prior work has aimed to reduce these gaps through test-time translation 27 ; 14 , model merging 35 ; 7 , and training-based methods 37 ; 26 ; 2 ; 13 ; 8 ; 16 . With the emergence of RLMs, recent work has further examined how multilingual inputs are processed within their reasoning traces. These analyses show that RLMs often produce English-centric reasoning traces even for non-English inputs, and that reasoning in English is associated with stronger performance 34 ; 28 ; 24 . 11 further show that understanding failures—i.e., failures to implicitly comprehend or translate the input into the trace language—are a major source of multilingual reasoning gaps in RLMs. To mitigate this, they propose Selective Translation, which uses an external detector to invoke translation when such understanding failures are predicted. Other recent methods aim to improve multilingual understanding jointly with reasoning ability during optimization 15 ; 5 . While effective, these approaches either rely on a separately trained detector, which can introduce additional latency, or remain bounded by the model’s own multilingual understanding, limiting gains on low-resource inputs. Luar addresses this gap by training the RLM itself to invoke an external translator when direct reasoning is unreliable, making translation an explicit action during reasoning. 3 Method Luar trains an RLM to selectively invoke translation when the input is likely to fall outside its language-understanding boundary. We first introduce an outcome-based surrogate for this boundary (§ 3.1 ), and then describe Luar ’s two-stage training procedure (§ 3.2 ). Figure 3 illustrates an overview of Luar . 3.1 Outcome-Based Boundary Estimation Okay, let’s see. As I read the question, some parts are difficult for me to fully understand. I’ll translate it into English to avoid any misunderstanding. Let me call the translator to get an English version of the question: { question in the original language } { ground-truth English reference translation } Now that I have the translated version, I can proceed with the reasoning step by step. Figure 2: Translator-call prefix prepended to the reasoning trace for translator-augmented reasoning. Figure 3: Overview of Luar . We first derive an outcome-based translation-usefulness label u(q) by comparing direct and translator-augmented reasoning outcomes. Stage I uses supervised fine-tuning to warm up both reasoning trajectories, and Stage II applies GRPO with a boundary-aware reward to learn when to invoke translation. Our goal is to identify cases where invoking translation is more reliable than directly reasoning over the original-language query. We therefore adopt an operational definition of the model’s language-understanding boundary: a query falls outside this boundary when access to an English translation substantially improves the model’s ability to solve it. We estimate this by sampling responses from two reasoning modes on the same query. For each multilingual query q , direct reasoning samples are generated by solving the original-language query without translation, while translator-augmented reasoning samples are generated by prepending an English reference translation to the reasoning trace using the translator-call prefix in Figure 2 . This intervention conditions the reasoning trace on the intended English meaning of the input while avoiding translation noise. We sample N/2 responses from each mode, with N=8 in our experiments. Let d_{\mathrm{acc}}(q) and t_{\mathrm{acc}}(q) be the fractions of correct responses under direct and translator-augmented reasoning, respectively. We define an outcome-based translation-usefulness label u(q)\in\{0,1\} as u(q)=\mathbb{I}\left[t_{\mathrm{acc}}(q)\geq d_{\mathrm{acc}}(q)+\delta\right] (1) where u(q)=1 means that q is estimated to fall outside the model’s language-understanding boundary, making translation useful for solving it. We use \delta=0.5 throughout our experiments, so that u(q)=1 reflects a substantial empirical benefit from translation. 3.2 Luar : Two-Stage Training Given this boundary estimate, Luar trains the model in two stages. Stage I extends the model’s action space by enabling it to produce translator-augmented reasoning traces. Stage II then optimizes this behavior with reinforcement learning, so that the model learns when translation should be invoked. Stage I: Translator-Call SFT. Stage I performs supervised fine-tuning (SFT) as a warm-up step before reinforcement learning. We sample responses from the base model under both direct and translator-augmented reasoning modes, and construct supervised targets from correct samples. For each query q , let \mathcal{Y}_{d}(q) and \mathcal{Y}_{t}(q) denote the sets of correct responses sampled under the two modes, respectively, where translator-augmented responses use the translator-call prefix in Figure 2 . We select the SFT mode m(q)\in\{d,t\} as m(q)=\begin{cases}t,&\text{if }t_{\mathrm{acc}}(q)>d_{\mathrm{acc}}(q),\\ d,&\text{otherwise}.\end{cases} (2) Since Stage I warms up exploration rather than defining the final policy, we use this softer rule instead of the stricter translator usefulness label u(q) in ( 1 ). This introduces potentially useful translator-augmented trajectories within the model’s sampling space, allowing Stage II to learn when to invoke translation through reinforcement learning. We then sample one correct response from the selected mode: y_{\mathrm{sft}}(q)\sim\mathrm{Uniform}\left(\mathcal{Y}_{m(q)}(q)\right), (3) removing queries for which the selected mode has no correct response. Finally, we fine-tune the model to predict y_{\mathrm{sft}}(q) given q . Stage II: Boundary-Aware GRPO. After Stage I, the model can generate both direct and translator-augmented responses. Stage II further optimizes this behavior through online reinforcement learning, so that the model learns to invoke translation selectively. To this end, we use GRPO 25 with a boundary-aware reward (Eq. ( 4 )) that jointly scores task correctness and translator-call decisions. For each query q , we sample a group of G responses \{y_{i}\}_{i=1}^{G} from the old policy \pi_{\theta_{\mathrm{old}}} . Unlike the boundary estimation in § 3.1 , this stage does not force a balanced number of direct and translator-augmented responses; each response is generated by the model’s current policy, which may decide whether to call the translator within its reasoning trace. When the model calls the translator, we provide the reference English translation of the input query as the translator response. This lets the model learn when translation is useful without confounding from translation quality. For each response y_{i} , let c(y_{i})\in\{0,1\} indicate whether the response gives the correct answer and conforms to the required output format, 2 2 2 See Appendix C for details on the format-checking criteria. and let t(y_{i})\in\{0,1\} indicate whether it calls the translator. Using the translation-usefulness label u(q) from ( 1 ), we define the boundary-aware reward as r(q,y_{i})=\begin{cases}1+\gamma,&c(y_{i})=1,\;t(y_{i})=1,\;u(q)=1,\\ 1,&c(y_{i})=1,\;t(y_{i})=0,\\ 1-\gamma,&c(y_{i})=1,\;t(y_{i})=1,\;u(q)=0,\\ 0,&c(y_{i})=0,\end{cases} (4) where \gamma=0.1 in our experiments. This reward prioritizes task correctness while encouraging translator calls only when translation is estimated to be useful. Incorrect or format-invalid responses receive zero reward, regardless of translator use. Among correct responses, direct reasoning receives the base reward, while translator calls receive a bonus when u(q)=1 and a penalty when u(q)=0 . 3 3 3 When all responses in the sampled group use the same mode, we omit the translator-use adjustment \gamma and set r(q,y_{i})=c(y_{i}) . We normalize the rewards within each sampled group to obtain group-relative advantages: \hat{A}_{i}=\frac{r(q,y_{i})-\mathrm{mean}(\{r(q,y_{j})\}_{j=1}^{G})}{\mathrm{std}(\{r(q,y_{j})\}_{j=1}^{G})}. (5) We then optimize the clipped token-level objective \begin{aligned} \mathcal{J}_{\mathrm{GRPO}}(\theta)&=\mathbb{E}_{\{y_{i}\}_{i=1}^{G}\sim\pi_{\theta_{\mathrm{old}}}(\cdot\mid q)}\Bigg[\frac{1}{\sum_{i=1}^{G}|y_{i}|}\sum_{i=1}^{G}\sum_{t=1}^{|y_{i}|}\\ &\quad\min\Big(\rho_{i,t}\hat{A}_{i},\mathrm{clip}(\rho_{i,t},1-\epsilon_{l},1+\epsilon_{h})\hat{A}_{i}\Big)\Bigg],\end{aligned} (6) where \rho_{i,t} is the token-level importance sampling ratio. Following DAPO-style training 36 , we minimize the token-level policy gradient loss with \epsilon_{l}=0.2 and \epsilon_{h}=0.3 , without a KL divergence loss. For both Stages I and II, we mask out the tokens inside and during loss computation for training stability, following 4 . 4 Experiments 4.1 Experimental Setup Models. We primarily evaluate Luar on Qwen3-4B and additionally on Qwen3-8B 32 . Datasets. We evaluate on two multilingual reasoning benchmarks. PolyMath 30 measures mathematical reasoning, and we use its low-, medium-, and high-difficulty subsets, each with 125 test examples per language. To evaluate out-of-domain generalization beyond mathematics, we also use STEM-related categories from MMLU-ProX-Lite 31 , which contains 257 test examples per language. Evaluation covers 10 typologically diverse languages. Following 10 , we group them into high-resource languages (English (en), Chinese (zh), Arabic (ar), Korean (ko)), mid-resource languages (Thai (th), Bengali (bn)), and low-resource languages (Swahili (sw), Telugu (te), Yoruba (yo), Zulu (zu)). For training, we use the training split of DeepScaleR 18 , which is an English mathematical problem solving dataset. After contamination filtering against PolyMath, we translate the remaining questions into Arabic, Thai, and Swahili using GPT-4.1 21 . We choose these languages to expose the policy to different scripts and resource conditions. Stage I uses 10K examples, and Stage II uses a disjoint set of 10K examples; in both stages, examples are evenly distributed across the three training languages. Appendix B provides further details on the datasets and training-data construction. Evaluation metrics. We report task accuracy for reasoning performance. To evaluate whether a method invokes translation when it is actually useful, we report mode-selection macro-F1 . For each query q , the reference label u(q) is the outcome-based usefulness label from § 3.1 , where u(q)=1 means translator-augmented reasoning is useful and u(q)=0 means it is not. The predicted label \hat{u}(q) is given by the method’s evaluation-time decision: \hat{u}(q)=1 if it calls the translator and \hat{u}(q)=0 otherwise. We compute macro-F1 over the two usefulness classes to account for class imbalance. For each method, the usefulness labels are estimated using the model being evaluated. Baselines. We compare against four groups of baselines. The first three groups test alternative ways of making translator-call decisions, while the last group tests whether training alone can close the accuracy gap without using an external translator at inference time. (1) Prompt-based methods: Self-Assessment prompts the model to solve the problem and then report whether the original input was understandable; if it reports understanding failure, inference is rerun with translation. Native-Tool-Use equips the model with a translator function through its default function-calling interface, allowing the model to decide whether to call translation when the input is difficult to interpret. (2) External detector-based methods: Selective Translation ( ST ) is an external detector-based baseline that uses a prober to decide whether each input should be translated ( 11 ) . We compare two variants that predict the same translator-usefulness label from different prober inputs: ST (qr) uses the final hidden state after the question and generated reasoning trace, following the original setup, whereas ST (q) uses the final hidden state after the question only, matching the decision timing of Luar . (3) SFT-based method: Boundary-SFT internalizes the same outcome-based boundary signal through supervised fine-tuning. For each query, it fine-tunes on correct direct or translator-augmented traces selected according to the strict translation-usefulness label u(q) . (4) Translator-free training methods: We compare with standard GRPO 25 , as well as Q-Align 37 and TAPO 5 , which we adapt for RLMs. We include these baselines in the accuracy comparison to cover three training-only alternatives: Q-Align improves language understanding through translation-aligned supervision, standard GRPO improves reasoning using the same RL optimization setup as Luar but with a correctness-only reward c(y) , and TAPO jointly trains language understanding and reasoning through translation-augmented policy optimization. We also report Full-Translation as a high-cost reference that translates every input. Further details are provided in Appendix E . Method PolyMath-Low PolyMath-Medium PolyMath-High MMLU-ProX-Lite Average (↑) Qwen3-4B Self-Assessment 41.10_{\pm 0.28} 56.44_{\pm 2.55} 52.93_{\pm 1.48} 47.83_{\pm 1.04} 49.57_{\pm 0.83} Native-Tool-Use 57.60_{\pm 1.90} 49.61_{\pm 0.19} 48.54_{\pm 2.36} 60.03_{\pm 0.48} 53.95_{\pm 1.13} ST (q) 63.25_{\pm 0.00} 52.59_{\pm 0.00} 53.58_{\pm 0.00} 48.38_{\pm 0.00} 54.45_{\pm 0.00} ST (qr) 79.26_{\pm 0.50} 64.59_{\pm 0.42} 58.24_{\pm 1.89} 65.44_{\pm 0.13} 66.88_{\pm 0.52} Boundary-SFT 73.14_{\pm 0.69} 62.22_{\pm 4.75} 66.85_{\pm 1.32} 65.21_{\pm 0.44} 66.85_{\pm 1.80} Luar \mathbf{87.64}_{\pm 0.33} \mathbf{75.22}_{\pm 1.84} \mathbf{73.74}_{\pm 1.11} \mathbf{76.45}_{\pm 1.02} \mathbf{78.26}_{\pm 0.42} Qwen3-8B Self-Assessment 43.48_{\pm 0.83} 54.28_{\pm 2.56} 48.60_{\pm 1.92} 48.04_{\pm 0.38} 48.60_{\pm 0.47} Native-Tool-Use 58.41_{\pm 3.53} 39.23_{\pm 0.35} 45.60_{\pm 2.03} 41.36_{\pm 0.59} 46.15_{\pm 0.67} ST (q) 64.01_{\pm 0.00} 56.18_{\pm 0.00} 60.41_{\pm 0.00} 46.80_{\pm 0.00} 56.85_{\pm 0.00} ST (qr) \mathbf{82.92}_{\pm 0.94} \mathbf{69.19}_{\pm 0.33} 60.73_{\pm 0.84} 61.13_{\pm 0.05} 68.49_{\pm 0.11} Boundary-SFT 69.20_{\pm 0.44} 65.90_{\pm 1.72} 60.54_{\pm 1.32} 59.47_{\pm 0.50} 63.78_{\pm 0.55} Luar 78.32_{\pm 0.11} 65.61_{\pm 1.29} \mathbf{69.74}_{\pm 1.34} \mathbf{68.18}_{\pm 0.55} \mathbf{70.46}_{\pm 0.44} Table 1: Mode-selection macro-F1 (%) for translator-call decisions. Numbers are mean ±std over seeds. Implementation details. We implement Luar using VerlTool 9 . Stage I uses supervised fine-tuning for 3 epochs with a learning rate of 1\times 10^{-5} and batch size 16. Stage II uses reinforcement learning for 150 steps with a learning rate of 2\times 10^{-6} , batch size 32, mini-batch size 16, group size 8, and rollout temperature 1.0. We use the same SFT and RL configurations for all corresponding training baselines. For evaluation, we average results over three random seeds and use the same decoding configuration across all methods. Appendix C and Appendix D provide additional training and evaluation details. 4.2 Main Results Luar effectively learns the language-understanding boundary. We first test whether Luar internalizes the outcome-based boundary signal as an effective mode-selection policy. Table 1 reports the mode-selection macro-F1, which measures whether each method selects the appropriate reasoning mode according to the estimated boundary labels. Overall, Luar achieves the best average macro-F1 for both models, showing that the boundary signal can be effectively internalized into the RLM policy. 4 4 4 Appendix G.1 shows that Luar maintains strong mode-selection performance when evaluated under different boundary thresholds \delta . Prompt-based methods show limited performance, and Boundary-SFT remains below Luar , suggesting that prompting or supervised fine-tuning alone is less effective for boundary-aware decisions. Compared with external detector-based selective translation, Luar achieves the best average macro-F1, outperforming ST (qr) on Qwen3-4B and remaining competitive on Qwen3-8B. Since macro-F1 alone does not reveal how translator calls are distributed across languages, we further examine translator-call rates across methods and languages. Figure 4: Translator call rate across languages and datasets. Luar rarely invokes translation for high-resource languages and instead concentrates translator calls on low-resource languages, while external selective translation tends to over-call the translator in broader settings. Luar learns a robust and targeted translator-call policy. Figure 4 shows how each method distributes translator calls across languages. Luar rarely invokes translation for high-resource languages, but frequently calls the translator for low-resource languages such as Swahili, where the base model performs poorly. This pattern also generalizes to unseen low-resource languages such as Yoruba and Zulu. In contrast, ST (qr) shows less stable call behavior: outside PolyMath-Low, it often calls the translator even for high-resource languages where translation is rarely needed; for Qwen3-8B on PolyMath-Medium, it calls the translator for 32% of English examples. Together, these results suggest that Luar learns a translator-call policy that is robust across evaluation settings and transfers to unseen low-resource languages. 4.3 Task Accuracy Method PolyMath-Low PolyMath-Medium PolyMath-High MMLU-ProX-Lite Average (Use%) Base 64.40_{\pm 0.16} 32.05_{\pm 0.24} 19.76_{\pm 0.08} 63.49_{\pm 0.49} 44.93_{\pm 0.14} (0.0%) Full-Translation 88.59_{\pm 0.44} 35.79_{\pm 1.09} 24.24_{\pm 0.79} 72.94_{\pm 0.31} 55.39_{\pm 0.05} (100.0%) Self-Assessment 63.52_{\pm 0.76} 36.00_{\pm 0.29} 23.95_{\pm 0.17} 64.36_{\pm 0.34} 46.96_{\pm 0.20} (15.0%) Native-Tool-Use 70.27_{\pm 0.65} 34.93_{\pm 0.36} 23.73_{\pm 0.23} 65.77_{\pm 0.35} 48.68_{\pm 0.16} (9.9%) ST (q) 72.32_{\pm 0.50} 32.85_{\pm 0.38} 20.80_{\pm 1.10} \mathbf{71.80}_{\pm 0.02} 49.44_{\pm 0.44} (21.7%) ST (qr) 79.36_{\pm 0.16} 35.47_{\pm 0.54} 24.32_{\pm 0.48} 71.57_{\pm 0.31} 52.68_{\pm 0.31} (35.5%) Boundary-SFT 74.20_{\pm 1.53} 31.56_{\pm 0.96} 20.68_{\pm 0.40} 63.60_{\pm 0.36} 47.51_{\pm 0.05} (14.7%) Q-Align 62.43_{\pm 0.28} 30.64_{\pm 0.40} 19.07_{\pm 0.65} 59.79_{\pm 0.25} 42.98_{\pm 0.12} (0.0%) GRPO 63.71_{\pm 0.88} 35.79_{\pm 0.60} 22.96_{\pm 0.65} 61.41_{\pm 0.20} 45.97_{\pm 0.16} (0.0%) TAPO 61.20_{\pm 0.21} 36.13_{\pm 0.18} 23.41_{\pm 0.68} 61.25_{\pm 0.12} 45.50_{\pm 0.19} (0.0%) Luar \mathbf{82.88}_{\pm 0.65} \mathbf{37.95}_{\pm 1.19} \mathbf{28.32}_{\pm 0.63} 68.43_{\pm 0.34} \mathbf{54.39}_{\pm 0.07} (20.3%) Table 2: Accuracy (%) across evaluation datasets on Qwen3-4B, reported as mean ±std over seeds. The Average column reports translator usage rate (%). Bold indicates the best result among non-full-translation methods. Luar improves multilingual reasoning performance. Table 2 reports final-answer accuracy on Qwen3-4B. The Average column reports the average accuracy across evaluation datasets, with the translator usage rate shown in parentheses. Full-Translation serves as a high-cost reference. Among methods that do not translate every input, Luar achieves the highest average accuracy, improving the base model from 44.93 to 54.39 while translating only 20.3% of examples. Luar ’s gains are especially strong on PolyMath, where it achieves the best accuracy across all three difficulty levels. On MMLU-ProX-Lite, Luar also improves over the base model, but remains below ST (q) and ST (qr) . However, this advantage for external detector-based methods comes with broader translator use, including unnecessary calls on high-resource languages (Figure 4 ). Translator-free training baselines show only limited improvements and remain below methods that can invoke translation, highlighting the promise of leveraging an external translator for multilingual reasoning in RLMs. Method Seen Unseen Avg. ar th sw en zh ko bn te yo zu Base 89.33 83.73 32.53 95.73 86.93 89.60 82.93 66.67 8.53 8.00 64.40 Full-Translation 91.20 90.40 84.80 95.73 92.80 90.93 91.20 81.87 78.40 88.53 88.59 Self-Assessment 90.67 82.40 30.67 94.93 88.00 89.33 82.67 66.93 5.07 4.53 63.52 Native-Tool-Use 89.87 83.47 40.53 94.13 88.80 89.60 83.73 70.13 29.33 33.07 70.27 ST (q) 89.07 83.47 67.20 95.47 87.47 89.87 83.47 69.60 20.80 36.80 72.32 ST (qr) 90.13 85.33 52.53 96.00 88.27 89.87 84.27 69.87 58.93 78.40 79.36 Boundary-SFT 90.00 83.20 54.80 94.80 83.60 84.40 79.60 69.60 30.80 71.20 74.20 Q-Align 87.73 84.00 38.40 95.73 86.93 84.00 77.33 63.20 2.67 4.27 62.43 GRPO 88.00 83.20 29.60 94.93 87.73 89.33 80.80 65.60 11.20 6.67 63.71 TAPO 83.20 81.07 39.20 95.47 85.87 84.53 72.00 61.33 3.73 5.60 61.20 Luar 89.60 85.07 89.60 95.73 89.07 86.67 81.60 68.00 56.27 87.20 82.88 Table 3: Per-language task accuracy (%) on PolyMath-Low with Qwen3-4B. Numbers are mean accuracy over seeds. Cell colors indicate changes from the Base row, clipped at \pm 20 points. Luar ’s accuracy gains are concentrated in languages where direct reasoning is weakest. Table 3 reports per-language accuracy on PolyMath-Low. The base Qwen3-4B model performs better on high-resource languages than on low-resource languages, with substantial degradation in the latter group. Translator-free training baselines improve the seen low-resource language Swahili, but transfer poorly to unseen low-resource languages such as Yoruba and Zulu. In contrast, Luar substantially improves these unseen low-resource languages, reaching 56.27 on Yoruba and 87.20 on Zulu, and achieves the best average accuracy among non-full-translation methods. These results further confirm the benefit of Luar : it improves accuracy precisely where direct language understanding is unreliable, while avoiding full translation. Additional per-language accuracy results are reported in Appendix G.3 . 4.4 Ablation Study We conduct an ablation study to analyze how the training objective shapes translator-call behavior. We compare Luar with four variants. First, Stage 1 only uses first-stage supervised training only. Second, GRPO with accuracy reward uses only the task success reward, r=c(y) , where c(y) denotes the accurate-and-format-correct indicator used in Eq. 4 . Third, GRPO with static penalty subtracts a fixed \gamma=0.1 penalty from successful translator-call responses, regardless of whether translation is useful. Finally, Luar (stage 1 strict) uses the strict usefulness label for Stage 1 target selection, instead of the softer rule t_{\mathrm{acc}}(q)>d_{\mathrm{acc}}(q) used by Luar . Figure 5: Validation translator-call rate of ablation methods during training. Method Accuracy (%) Macro-F1 (%) Stage 1 only 50.46_{\pm 0.16} 71.85_{\pm 0.39} + GRPO, accuracy reward 53.14_{\pm 0.08} 70.79_{\pm 0.79} + GRPO, static penalty ( \gamma=0.1 ) 50.33_{\pm 0.12} 70.55_{\pm 0.17} Luar (Ours) \mathbf{54.39}_{\pm 0.07} \mathbf{78.26}_{\pm 0.42} Luar (stage 1 strict) 53.66_{\pm 0.06} 74.86_{\pm 0.59} Table 4: Ablation of RL training and reward design. Accuracy and mode-selection macro-F1 are averaged over evaluation datasets and seeds. Luar ’s boundary-aware reward is necessary for selective translator-call learning. Figure 5 shows that alternative objectives lead to undesirable translator-call dynamics during training. The accuracy-reward GRPO baseline rapidly increases the validation translator-call rate, whereas the static penalty baseline suppresses translator calls throughout training. In contrast, Luar avoids both extremes by rewarding translator calls only when they are supported by the outcome-based boundary signal. Table 4 further shows that this selective call behavior yields better final performance: Luar improves both task accuracy and mode-selection macro-F1 over Stage 1 only, while the alternative objectives either trade off accuracy against macro-F1 or degrade both. The strict Stage 1 variant also underperforms full Luar , suggesting that the softer Stage 1 target selection helps expose translator-call behavior before refinement with the stricter boundary-aware reward. 4.5 Efficiency Method Latency (s) Input Tokens Output Tokens Translator Use (%) ST (qr) 14.09 687.0 10794.4 30.0 Base 11.77 288.9 9251.0 0.0 Luar 9.32 505.6 6652.6 20.5 Table 5: Overall efficiency statistics of Qwen3-4B from a single run, averaged across all evaluation datasets. To quantify the inference efficiency of our method, we compare Luar against the Base model and external selective translation, ST (qr) , using the Qwen3-4B backbone across all evaluation datasets. As shown in Table 5 , Luar is more efficient than ST (qr) . Notably, even though it uses translation for 20.5% of examples, Luar also achieves lower latency than the Base model by producing shorter reasoning traces. These results highlight the inference efficiency of Luar , which internalizes the translation decision during reasoning. 5 Conclusion We introduced Luar , a Language Understanding Boundary-aware Reinforcement Learning framework for multilingual reasoning. Luar improves reasoning performance by selectively invoking translation when direct understanding is unreliable and generalizes to unseen languages across multilingual reasoning benchmarks. These findings highlight language-understanding boundaries as a useful perspective for building robust multilingual reasoning models. Future work could extend boundary-aware policy learning to broader multilingual tasks and abstention settings. Limitations Although our evaluation includes typologically diverse languages across different resource conditions, it does not cover the full diversity of language families. Evaluating Luar on a broader set of languages would provide a more comprehensive understanding of its cross-lingual generalization. We also evaluate Luar on mathematical and STEM reasoning benchmarks, which differ in domain and question format. However, multilingual reasoning spans a wider range of tasks beyond these settings, including commonsense and knowledge-intensive reasoning. Extending Luar to more diverse multilingual reasoning tasks would further test the robustness and generality of boundary-aware translation-use policies. Ethical Considerations In this work, we use DeepScaleR 18 for training, and Polymath 30 and MMLU-ProX-Lite 31 for evaluation. These datasets are licensed under MIT, Apache 2.0, and MIT, respectively. We conduct experiments with Qwen3-4B/8B 32 and gpt-oss-20b/120b 19 , all of which are licensed under Apache 2.0. All datasets and models are used strictly for research purposes and in accordance with their intended use. We used ChatGPT and Codex for language polishing, grammar correction, and programming assistance. All ideas, contributions, and research decisions are solely those of the authors. Acknowledgments This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2025 (Project Name: Development of an AI-Based Korean Diagnostic System for Efficient Korean Speaking Learning by Foreigners, Project Number: RS-2025-02413038, Contribution Rate: 45%); by the IITP(Institute of Information & Coummunications Technology Planning & Evaluation)-ITRC(Information Technology Research Center) grant funded by the Korea government(Ministry of Science and ICT)(IITP-2026-RS-2024-00437866, Contribution Rate: 45%); and by the Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2019-II191906, Artificial Intelligence Graduate School Program(POSTECH), Contribution Rate: 10%). Swahili question Dada wa Doubtfire wanaendesha wakielekea nyumbani na watoto 7 wa paka waliochukua kutoka makao ya wanyama katika eneo lao kisha mama yao akawapigia simu kuwa paka wao wawili wa nyumba wamezaa watoto. Anasema kuwa Patchy, paka wa kwanza, ana watoto ambao idadi yao ni mara tatu ya watoto wa paka waliotwaa, huku Trixie, paka mwingine, ana watoto 12. Familia ya Doubtfire sasa ina watoto wangapi wa paka? (Translation: The Doubtfire sisters are driving home with 7 kittens adopted from the local animal shelter when their mother calls to inform them that their two house cats have just had kittens. She says that Patchy, the first cat, has had thrice the number of adopted kittens, while Trixie, the other cat, has had 12. How many kittens does the Doubtfire family now have?) Reasoning trace We need to parse the problem: It’s in Swahili? Let’s translate: “Dada wa Doubtfire wanaendesha wakielekea nyumb…..So: “Doubtfire’s sister runs a home for stray cats (?) …The phrase “wakielekea nyumbani” might be “to bring home” or “to drive home”?… Actually “wakielekea nyumbani” might mean “to bring home” or “to drive home”…But we need to decide which interpretation is correct…. But the question might ask for “how many cat children does the Doubtfire family have?” That includes the 7 kittens from the shelter and the new ones. So answer 55. But maybe the 7 cats are adult cats, not kittens. But the phrase “watoto 7 wa paka” is ambiguous. But I’d lean that it’s 7 kittens.Thus answer: \boxed{55} . Let’s provide final answer: 55. Final response Patchy has 3\times 12=36 kittens, and Trixie has 12 kittens. The family already had 7 kitten-children from the shelter. 7 + 36 + 12 = 55. \boxed{55} ✗ Table 6: An example from gpt-oss-20b on PolyMath-Low where explicit uncertainty about an understanding failure appears in the reasoning trace but is omitted from the final response. Appendix A Explicit Uncertainty in Understanding Failures Reasoning language models (RLMs) often exhibit difficulty in understanding queries written in languages with which they are less familiar ( 11 ) . A representative example is shown in Table 6 . In this example, the model’s reasoning trace explicitly expresses uncertainty about the Swahili wording and considers alternative interpretations, indicating difficulty in understanding the query. This uncertainty leads the model to guess a key condition in the problem: instead of using the fact that Patchy has three times the number of adopted kittens, i.e., 3\times 7 , the model incorrectly computes Patchy’s kittens as 3\times 12 . As a result, the model produces an incorrect answer due to an understanding failure. However, this uncertainty is not reflected in the final response. Although the reasoning trace reveals that the model may not have properly understood the question, the final response presents the solution confidently without acknowledging this uncertainty. This is problematic because users typically observe only the final response, and therefore may not be aware that the answer is based on an uncertain or potentially incorrect understanding of the query. To better understand how often models explicitly express such uncertainty during response generation, we measure the explicit uncertainty rate separately in the reasoning trace and in the final response. We measure explicit uncertainty in understanding-failure cases through a two-stage evaluator pipeline. Stage 1. The first stage identifies examples in which the model’s final response reflects a critical misunderstanding of the original question. We conduct this analysis on PolyMath-Low ( 30 ) , using the same languages as in the main experiments. For each of Qwen3-4B, Qwen3-8B 32 , gpt-oss-20b, and gpt-oss-120b 19 , we inspect only examples whose final answer is incorrect. Given the original question and the final response generated after the reasoning trace, we use gpt-5.5 5 5 5 https://developers.openai.com/api/docs/models/gpt-5.5 ( reasoning_effort=high ) to determine whether the final response critically misunderstands the conditions specified in the question. Table 7 summarizes the results of Stage 1. The evaluator prompt for Stage 1 is shown below. Understanding-Failure Detection Prompt You are an evaluator of multilingual math question understanding. Input: - question: {question} - response: {response} Your task is to determine whether the response shows a critical misunderstanding of the question. Return "fail": true only if the response shows a critical misunderstanding of the question that affects the solution. Critical misunderstandings include: - solving a different problem - misunderstanding key quantities, objects, units, conditions, variables, or comparison targets - ignoring or mistranslating multilingual text in a way that changes the meaning Return "fail": false if the response understood the intended question well enough, even if it has minor arithmetic mistakes, wording issues, or an incomplete explanation. Output format: Return only a JSON object with a single boolean field: {"fail": true} or {"fail": false} Model Incorrect Understanding failure Failure rate Qwen3-4B 433 358 82.7 Qwen3-8B 346 292 84.4 gpt-oss-20b 157 92 58.6 gpt-oss-120b 142 59 41.5 Table 7: Stage 1 statistics for evaluating explicit uncertainty in understanding-failure cases on PolyMath-Low (1250 examples). Stage 2. Among the examples identified in Stage 1, the second stage computes the explicit uncertainty rate. We define explicit uncertainty as an explicit statement by the model that the wording, language, translation, or meaning of the question is confusing, ambiguous, difficult to parse, or uncertain. This definition excludes ordinary step-by-step setup phrases such as “let me understand the problem,” wrong restatements, arithmetic errors, and incorrect answers, unless the model explicitly expresses uncertainty about the meaning of the question. To examine where this phenomenon appears during response generation, we apply this check separately to the reasoning trace and to the final response. Because this stage requires a direct surface-level judgment about whether such an explicit statement appears, we use gpt-oss-120b as the evaluator. The evaluator prompts used in Stage 2 for reasoning traces and final responses are shown below. Reasoning Trace Verbalized Confusion Detection Prompt You are a careful evaluator of model behavior. You are given: 1. The original question. 2. The model’s reasoning trace while answering that question. Decide whether the model text explicitly verbalizes difficulty understanding the question. Return true only when the model itself says the wording, language, translation, or meaning of the question is confusing, ambiguous, hard to parse, or uncertain. Examples include statements like "the wording is confusing", "I need to parse this more carefully", "does this mean X or Y?", or "I am not sure what the question is asking". Return false for normal step-by-step reasoning, arithmetic mistakes, wrong answers, wrong restatements, or phrases like "let me understand/break down the problem" unless the model explicitly expresses confusion or uncertainty about the question’s meaning. Return exactly this JSON object: { "explicit_understanding_failure": true or false } Original question: question Reasoning trace to inspect: model_text Final Response Verbalized Confusion Detection Prompt You are a careful evaluator of model behavior. You are given: 1. The original question. 2. The model’s final prediction for that question. Decide whether the model text explicitly verbalizes difficulty understanding the question. Return true only when the model itself says the wording, language, translation, or meaning of the question is confusing, ambiguous, hard to parse, or uncertain. Examples include statements like "the wording is confusing", "I need to parse this more carefully", "does this mean X or Y?", or "I am not sure what the question is asking". Return false for normal step-by-step reasoning, arithmetic mistakes, wrong answers, wrong restatements, or phrases like "let me understand/break down the problem" unless the model explicitly expresses confusion or uncertainty about the question’s meaning. Return only a valid JSON object. Do not include Markdown, code fences, or extra text. Return exactly this JSON object: { "explicit_understanding_failure": true or false } Original question: question Final prediction to inspect: model_text Computing the explicit uncertainty rate. Let \mathcal{F}_{m} be the set of examples for model m whose final answer is incorrect and whose final response is judged as an understanding failure by the first-stage evaluator. Let u^{\mathrm{trace}}_{m,i} and u^{\mathrm{final}}_{m,i} indicate whether the second-stage evaluator finds explicit uncertainty in the reasoning trace and final response, respectively. We compute the explicit uncertainty rates as \mathrm{Rate}^{\mathrm{trace}}_{m}=\frac{1}{|\mathcal{F}_{m}|}\sum_{i\in\mathcal{F}_{m}}u^{\mathrm{trace}}_{m,i}, and \mathrm{Rate}^{\mathrm{final}}_{m}=\frac{1}{|\mathcal{F}_{m}|}\sum_{i\in\mathcal{F}_{m}}u^{\mathrm{final}}_{m,i}. These quantities correspond to the explicit uncertainty rates in the reasoning trace and final response, respectively, as reported in Figure 1 . Appendix B Dataset and Training Data Construction Evaluation datasets. We evaluate on PolyMath 30 and MMLU-ProX-Lite 31 to cover both in-domain mathematical reasoning and out-of-domain STEM reasoning. PolyMath is a multilingual mathematical reasoning benchmark that covers multiple difficulty levels. We use the Low, Medium, and High splits: Low consists of K–12 style mathematical word problems, Medium contains exam-style and entry-level competition problems, and High focuses on more difficult competition problems. Each split provides parallel questions across languages, allowing direct comparison of multilingual reasoning performance. For PolyMath, we additionally include Yoruba and Zulu by translating the English questions with GPT-4.1, using the dataset translation prompt described below. MMLU-ProX-Lite is a compact version of MMLU-ProX with parallel multiple-choice questions across languages. We use its STEM-related subjects ( math, physics, chemistry, computer science, and engineering ) as an out-of-domain reasoning evaluation beyond the mathematical training data. For both benchmarks, we follow the evaluation prompt format of 11 . Training source data. For training, we use the train split of agentica-org/DeepScaleR-Preview-Dataset . Since DeepScaleR 18 is a mathematical reasoning dataset and PolyMath is also used for evaluation, we apply n-gram contamination filtering before sampling training examples. We normalize every English question in all PolyMath difficulty splits (Low, Medium, High, and Top) and build a character 13-gram index. We then compute the character n-gram Jaccard similarity between each DeepScaleR problem and each PolyMath question. A DeepScaleR problem is removed as contaminated if it has Jaccard similarity of at least 0.8 with a PolyMath question and at least 10 overlapping character 13-grams. Training splits and languages. After filtering, we sample disjoint 10K-example sets for Stage 1 SFT and Stage 2 RL. The training languages are Arabic (ar), Thai (th), and Swahili (sw), and each split is balanced across these languages with an approximately 1{:}1{:}1 ratio. We reserve 200 validation examples for model selection, also sampled in a language-balanced way, and use the remaining examples for training. Translation for training data. We use GPT-4.1 21 with decoding temperature 0 to translate English DeepScaleR questions into the training languages. To filter obvious translation failures, we tokenize both the translated question and the English source question with the Qwen3 32 tokenizer and discard a translated question if its tokenized length is more than 15\times the tokenized length of the English source. Dataset Translation Prompt Translate the following mathematical question enclosed within and into {language_name}. The text may contain mathematical notation and LaTeX formatting. You must strictly preserve: - All LaTeX math and commands EXACTLY as written, including inline math $...$ , display math $$...$$ , \(...\) , \[...\] , and any \begin{...}...\end{...} environments. - All mathematical symbols, variables, numbers, operators, and equation labels. Provide only the translated problem without any additional explanation. Wrap the translated output with and tags. {problem} Appendix C Training Details Stage 1 SFT. Stage 1 uses AdamW 17 with learning rate 1\times 10^{-5} and batch size 16 . The maximum prompt length is 2,560 tokens and the maximum response length is 12,288 tokens. We train for 3 epochs, evaluate every 200 steps, and select the checkpoint with the best validation accuracy-and-format reward, which is 1 only when the answer is accurate and the required output format is correct. Stage 2 RL. Stage 2 uses group size G=8 , AdamW with learning rate 2\times 10^{-6} , batch size 32 , mini-batch size 16 , rollout temperature 1.0 , and top- p=1.0 . The maximum prompt length is 2,560 tokens and the maximum response length is 12,288 tokens. We use the asymmetric clipping parameters \epsilon_{l}=0.2 and \epsilon_{h}=0.3 . We train for 150 steps, evaluate every 10 steps, and select the checkpoint with the best validation accuracy-and-format reward. We use this criterion for all training-based methods because it applies uniformly to methods with and without translator decisions, whereas mode-selection macro-F1 is defined only for methods that make a translator-call decision. Training cost. Training Luar on two NVIDIA H200 GPUs takes approximately 5 hours for Stage 1 SFT and 13 hours for Stage 2 RL on Qwen3-4B, and 8 hours for Stage 1 SFT and 19 hours for Stage 2 RL on Qwen3-8B. Format correctness. During training, the correctness indicator c(y) requires both answer correctness and format correctness. A response is format-correct only if it contains exactly one tag and one tag, appears before , and the output starts with . The final solution block after must contain an extractable \boxed{...} answer. If the response uses ... , the translator call must be well formed, must not be nested, and its stripped content must contain at least five characters. The tags , , , and must not appear in the final solution block. For English queries, any translator tag makes the response format-invalid. Setting Value Optimizer AdamW Learning rate 1\times 10^{-5} Batch size 16 Max prompt length 2,560 Max response length 12,288 Epochs 3 Table 8: SFT hyperparameters for Luar Stage 1, Boundary-SFT , Q-Align, and TAPO Stage 1. Sequence lengths are measured in tokens. Setting Value Optimizer AdamW Learning rate 2\times 10^{-6} Batch size 32 PPO mini-batch size 16 Group size 8 Max prompt length 2,560 Max response length 12,288 Rollout temperature 1.0 Rollout top- p 1.0 Clipping parameters \epsilon_{l}=0.2,\epsilon_{h}=0.3 Training steps 150 Table 9: RL hyperparameters for Luar Stage 2, TAPO Stage 2, and Naive GRPO. Sequence lengths are measured in tokens. Setting Value Methods ST (q) and ST (qr) Backbone Frozen RLM Features Last-token final-layer hidden state Classifier hidden size \{0,32,128,512,d/2,d\} Learning rate \{10^{-3},10^{-4},10^{-5}\} Optimizer AdamW Weight decay 0.01 Batch size 64 Epochs 10 Loss Class-weighted cross-entropy Model selection Best validation macro-F1 Table 10: Training hyperparameters for the ST prober baselines. Here, d is the hidden-state dimension and hidden size 0 denotes a linear probe. Appendix D Evaluation Details Decoding parameters. We report metrics averaged over three seeds sampled with temperature 0.6, top-p = 0.95, top-k = 50, maximum prompt length 2,560, and maximum response length 12,288. Translation at evaluation time. When evaluation requires translation, we use GPT-4.1 21 with the tool-calling translation prompt shown below. This applies to both external translation baselines and translator calls made by Luar . We implement translator-call execution with VerlTool 9 : when the model emits a well-formed translator_call span, the environment appends the GPT-4.1 translation as a translator_response and continues decoding from the updated context. In contrast to training-time translator-augmented rollouts, evaluation uses GPT-based translations rather than reference English questions. Tool-Calling Translation Prompt Translate the following text enclosed within and into English. The text may contain mathematical notation and LaTeX formatting. In this case, you must strictly preserve: - All LaTeX math and commands EXACTLY as written. - All mathematical symbols, variables, numbers, operators, and equation labels. Provide only the translated text without any additional explanation. Wrap the translated text with and tags. {input_text} Answer extraction and verification. We follow 11 for answer extraction and verification. For PolyMath 30 , we extract answers from the \boxed{} pattern and verify them against the gold answers using Math-Verify. 6 6 6 https://github.com/huggingface/Math-Verify For MMLU-ProX-Lite 31 , we use a regular-expression matcher to extract the final multiple-choice answer and compare it with the gold option. Appendix E Baseline Implementation Details ST (q) and ST (qr) . We construct prober labels following the translator-usefulness estimation procedure in § 3.1 . A prober is a lightweight classifier trained on frozen hidden representations of the backbone model to predict whether a query is translator-useful. Specifically, for each DeepScaleR training question, we construct the translator usefulness label following the procedure in § 3.1 . Following 11 , we train the prober with class-weighted cross-entropy loss and use their grid-search protocol over learning rates \{10^{-3},10^{-4},10^{-5}\} and hidden-layer sizes \{0,32,128,512,d/2,d\} , where d is the hidden-state dimensionality and 0 corresponds to a linear probe. ST (q) extracts the final-layer hidden state at the last token of the question, while ST (qr) extracts the final-layer hidden state at the last token of the reasoning trace. In both cases, the backbone language model is frozen and only the prober is trained. The prober search space and training hyperparameters are summarized in Table 10 . Boundary-SFT . Boundary-SFT uses the same rollout-based translator-usefulness labels as Luar , constructed as described in § 3.1 . For each question, we select a translator-augmented target only when the translator-augmented accuracy is at least 0.5 higher than the direct accuracy; otherwise, we select a direct target. We retain only sampled responses that give the correct answer and conform to the required output format. The SFT hyperparameters are given in Table 8 . Q-Align . Q-Align 37 improves multilingual reasoning by training the model to translate non-English reasoning questions into English, thereby aligning non-English questions with their English counterparts and enabling transfer from English reasoning ability. Since our backbone is already a post-trained RLM, we avoid directly optimizing a separate question-translation objective, which may disrupt its learned reasoning behavior by encouraging translation-only outputs. Instead, we adapt Q-Align by inserting the gold English translation into the reasoning trace and training the model to solve the problem conditioned on it. We construct targets from the same sampled DeepScaleR training questions used for Luar . During sampling, we force the model to condition on the English translation of the problem before solving, using the prefix below. We retain only sampled responses that are answer-correct and format-valid. The sampling and SFT hyperparameters are summarized in Table 8 . Q-Align Forced Prefix Okay, let me first translate the question into English to make sure I understand it correctly. English translation: {prompt_text_in_english} Now that the question is clear in English, I will solve the problem step by step and derive the final answer. TAPO . TAPO 5 was originally proposed for instruction-tuned models, where the model is trained to follow an explicit understand-then-reason format: first translate the problem into English, and then solve it based on the translation. Since our backbone is already a post-trained RLM, directly imposing the original pipeline may disrupt its learned reasoning format and behavior. We therefore adapt TAPO by preserving its explicit translation-augmented reasoning structure, while introducing it into the RLM through a two-stage training pipeline. We use the original TAPO system prompt for both Stage 1 sampling and Stage 2 RL. Stage 1 transfers the TAPO reasoning format to the RLM: we sample DeepScaleR responses under the TAPO system prompt with a forced English-translation prefix, so that the resulting reasoning traces begin with an explicit English translation block before continuing to reasoning and the final solution. Only answer-correct and format-valid sampled responses are retained for SFT. Stage 2 then optimizes this behavior using the TAPO reward and step-level advantage formulation, which decouples credit assignment for translation and reasoning tokens. We use the same RL hyperparameters as Luar , except for the reward and advantage formulation. We use ChrF++ 23 score 7 7 7 We compute ChrF++ using the Hugging Face evaluate implementation: https://github.com/huggingface/evaluate/blob/main/metrics/chrf/README.md , with word_order=2 and lowercase=True . as the translation reward and set the advantage interpolation coefficient \alpha to 0.25 , following the original TAPO setup. The SFT and RL hyperparameters are listed in Tables 8 and 9 . TAPO Forced Translation Prefix {prompt_text_in_english} Naive GRPO. The Naive GRPO 25 baseline is trained on the same RL data as Luar and uses the same RL hyperparameters, as shown in Table 9 . The only difference is the reward: Naive GRPO uses an accuracy-and-format reward ( c(y) ) without Luar ’s translator-usefulness reward signal. Appendix F Prompt Formats This appendix lists the prompt and tool formats used for translator-augmented reasoning and prompting-based translator-calling baselines. F.1 Translator-Augmented Reasoning Translator-Augmented Reasoning Prefix Okay, let’s see. As I read the question, some parts are difficult for me to fully understand. I’ll translate it into English to avoid any misunderstanding. Let me call the translator to get an English version of the question: {prompt_text_in_original_language} prompt_text_in_english Now that I have the translated version, I can proceed with the reasoning step by step. F.2 Self-Assessment Baseline Self-Assessment Baseline Prompt You are a careful and reliable reasoning assistant. Solve the given question step by step. After solving the problem, assess whether you confidently understood the question itself in its original language. This judgment should reflect whether the question was understandable, not whether your answer is correct. If the question was written in an unfamiliar language, difficult to interpret, or if you are uncertain whether you understood it correctly, choose understandable: no. Otherwise choose understandable: yes. At the very end of your output, format your response exactly as follows: Answer: Understandable: yes/no F.3 Native Tool Use Baseline Native Tool Use Baseline Tool Description Tool name: translate_query_to_english Description: Translate the input query into English when necessary for accurate understanding. You may call this tool if the query is in a language you are not confident in interpreting. Avoid using it when the query is already in English or when its meaning is clear. Parameter: input_query: The input query to translate into English. Appendix G Additional Experimental Results G.1 Robustness to Translation-Usefulness Thresholds Figure 6: Mode-selection macro-F1 under different boundary thresholds on Qwen3-4B. Figure 7: Mode-selection macro-F1 under different boundary thresholds on Qwen3-8B. Our main experiments define the translation-usefulness label with \delta=0.5 in Eq. 1 for evaluation. To examine whether Luar ’s mode-selection advantage holds beyond this particular evaluation definition, we additionally evaluate macro-F1 under alternative boundary definitions. Specifically, we vary \delta\in\{0.25,0.5,0.75,1.0\} , and further consider a stricter setting where a query is labeled translator-useful only when t_{\mathrm{acc}}(q)=1.0 and d_{\mathrm{acc}}(q)=0.0 . Figure 6 and Figure 7 show that Luar maintains strong mode-selection performance across these alternative boundary definitions. In most settings, Luar achieves the best or highly competitive macro-F1 compared with prompt-based, external detector-based, and SFT-based baselines. The advantage is especially clear under stricter thresholds, where positive examples correspond to cases in which translation yields a large empirical improvement over direct reasoning. These results show that Luar remains effective not only under the main \delta=0.5 definition, but also when translation usefulness is evaluated with more relaxed or stricter criteria. G.2 Decision Diagnostics Figure 8: Precision and recall for translator-call decisions on Qwen3-4B. Luar improves decision quality while keeping translator use selective across datasets. Figure 9: Precision and recall for translator-call decisions on Qwen3-8B. ST (qr) attains high recall through substantially higher translator usage than Luar , but this comes with lower precision. G.3 Per-Language Results Method Seen Unseen Avg. ar th sw en zh ko bn te yo zu Qwen3-4B 89.33 83.73 32.53 95.73 86.93 89.60 82.93 66.67 8.53 8.00 64.40 Full-Translation 91.20 90.40 84.80 95.73 92.80 90.93 91.20 81.87 78.40 88.53 88.59 Self-Assessment 90.67 82.40 30.67 94.93 88.00 89.33 82.67 66.93 5.07 4.53 63.52 Native-Tool-Use 89.87 83.47 40.53 94.13 88.80 89.60 83.73 70.13 29.33 33.07 70.27 ST (q) 89.07 83.47 67.20 95.47 87.47 89.87 83.47 69.60 20.80 36.80 72.32 ST (qr) 90.13 85.33 52.53 96.00 88.27 89.87 84.27 69.87 58.93 78.40 79.36 Boundary-SFT 90.00 83.20 54.80 94.80 83.60 84.40 79.60 69.60 30.80 71.20 74.20 Q-Align 87.73 84.00 38.40 95.73 86.93 84.00 77.33 63.20 2.67 4.27 62.43 GRPO 88.00 83.20 29.60 94.93 87.73 89.33 80.80 65.60 11.20 6.67 63.71 TAPO 83.20 81.07 39.20 95.47 85.87 84.53 72.00 61.33 3.73 5.60 61.20 Luar 89.60 85.07 89.60 95.73 89.07 86.67 81.60 68.00 56.27 87.20 82.88 Qwen3-8B 90.40 89.87 62.67 96.53 89.33 91.47 89.33 77.60 13.60 15.73 71.65 Full-Translation 92.00 92.53 88.53 96.53 91.47 90.67 91.73 85.33 78.67 89.33 89.68 Self-Assessment 93.60 91.73 63.47 96.53 89.60 91.20 88.27 76.27 10.40 17.33 71.84 Native-Tool-Use 63.47 61.07 62.13 85.07 82.67 77.07 67.20 60.53 65.33 78.40 70.29 ST (q) 90.67 90.13 68.00 96.00 89.07 89.60 88.00 77.33 20.00 60.53 76.93 ST (qr) 90.40 90.67 78.67 96.27 89.87 90.40 88.53 81.33 63.47 69.07 83.87 Boundary-SFT 89.60 89.87 63.73 96.80 88.27 89.60 86.13 76.27 34.13 50.40 76.48 Q-Align 89.33 89.33 64.00 97.33 87.73 88.27 81.60 74.13 7.47 8.00 68.72 GRPO 90.40 90.67 62.67 96.53 89.60 91.20 86.40 75.20 7.47 13.87 70.40 TAPO 83.73 82.67 38.93 95.47 84.80 84.80 71.73 60.80 4.80 3.47 61.12 Luar 90.67 90.13 76.27 96.80 87.20 89.33 86.40 76.80 40.27 69.87 80.37 Table 11: Per-language task accuracy (%) on PolyMath-Low. Numbers are mean accuracy over seeds. Cell colors indicate changes from the Base row, clipped at \pm 20 points. Method Seen Unseen Avg. ar th sw en zh ko bn te yo zu Qwen3-4B 37.87 35.20 25.33 37.07 36.80 36.80 35.20 35.20 20.80 20.27 32.05 Full-Translation 36.80 36.80 35.47 37.07 37.87 36.27 34.40 33.60 33.60 36.00 35.79 Self-Assessment 41.07 37.60 35.20 39.20 43.20 42.40 39.73 37.87 20.80 22.93 36.00 Native-Tool-Use 39.73 37.87 29.60 40.00 40.80 41.07 37.07 34.40 24.00 24.80 34.93 ST (q) 36.27 35.47 28.80 36.27 37.60 38.93 35.20 34.40 21.07 24.53 32.85 ST (qr) 38.13 35.47 33.87 37.87 39.47 38.40 37.60 35.73 27.47 30.67 35.47 Boundary-SFT 32.80 32.00 31.60 34.40 32.80 35.60 35.20 30.80 21.60 28.80 31.56 Q-Align 34.93 33.60 29.87 34.40 35.47 34.67 33.07 33.60 18.40 18.40 30.64 GRPO 38.93 37.07 27.47 42.67 42.40 41.87 39.20 39.73 25.33 23.20 35.79 TAPO 40.00 36.80 37.60 42.67 41.07 41.07 38.93 38.93 21.07 23.20 36.13 Luar 40.53 38.93 37.87 40.27 41.33 40.27 40.27 34.93 27.47 37.60 37.95 Qwen3-8B 36.80 36.00 29.33 34.93 34.40 36.27 33.87 33.87 21.87 23.73 32.11 Full-Translation 36.53 33.60 36.53 34.93 36.80 36.80 35.20 34.13 34.93 35.73 35.52 Self-Assessment 41.33 40.00 35.73 38.40 41.33 43.20 38.67 38.93 21.60 27.73 36.69 Native-Tool-Use 22.13 21.07 27.47 31.47 28.53 16.53 22.67 29.33 25.07 24.80 24.91 ST (q) 37.87 36.27 30.13 36.27 35.73 36.27 33.07 34.13 26.13 29.87 33.57 ST (qr) 40.53 36.80 36.53 37.60 38.13 39.20 36.00 35.20 33.33 34.40 36.77 Boundary-SFT 37.33 34.67 32.80 35.47 35.20 36.80 35.20 31.73 25.07 30.93 33.52 Q-Align 34.93 33.60 33.60 34.93 34.40 33.07 33.60 30.40 18.93 20.53 30.80 GRPO 39.47 38.67 32.27 40.80 40.53 43.20 37.87 37.60 24.27 28.53 36.32 TAPO 38.40 36.80 31.20 41.07 37.07 38.93 38.67 34.40 17.60 20.80 33.49 Luar 42.40 40.53 37.60 44.00 43.20 41.60 39.20 40.27 25.87 32.27 38.69 Table 12: Per-language task accuracy (%) on PolyMath-Medium. Numbers are mean accuracy over seeds. Cell colors indicate changes from the Base row, clipped at \pm 20 points. Method Seen Unseen Avg. ar th sw en zh ko bn te yo zu Qwen3-4B 24.00 24.00 15.20 23.20 22.67 25.33 24.27 17.33 9.87 11.73 19.76 Full-Translation 25.60 24.80 25.33 23.20 21.60 24.00 24.27 22.13 25.33 26.13 24.24 Self-Assessment 31.73 26.40 20.53 28.53 28.00 29.87 26.93 26.67 9.33 11.47 23.95 Native-Tool-Use 28.80 24.80 17.33 28.27 29.87 29.87 25.33 25.07 13.33 14.67 23.73 ST (q) 26.40 24.80 17.07 22.93 24.00 25.33 23.20 20.27 11.73 12.27 20.80 ST (qr) 28.00 25.60 22.67 24.53 27.20 26.40 25.87 21.87 19.73 21.33 24.32 Boundary-SFT 24.40 22.40 20.00 22.80 17.60 23.20 25.60 21.20 11.60 18.00 20.68 Q-Align 22.93 21.60 18.40 20.80 23.20 21.33 22.40 21.07 9.87 9.07 19.07 GRPO 28.53 24.80 17.60 29.33 25.87 30.67 26.93 21.33 12.80 11.73 22.96 TAPO 28.00 26.40 21.60 28.53 30.13 26.40 26.67 28.00 9.33 9.07 23.41 Luar 30.67 29.33 32.53 30.93 29.07 32.27 27.20 25.07 18.40 27.73 28.32 Qwen3-8B 23.20 21.87 18.40 21.60 21.07 21.87 23.20 18.67 9.33 10.40 18.96 Full-Translation 23.73 23.47 22.93 21.60 24.53 24.27 23.20 23.20 25.33 25.60 23.79 Self-Assessment 29.07 28.80 24.27 25.07 30.93 29.60 30.67 29.07 12.80 14.13 25.44 Native-Tool-Use 20.27 18.67 20.27 21.60 24.00 14.93 19.73 23.20 20.80 18.67 20.21 ST (q) 22.13 24.27 20.80 22.40 20.80 22.93 23.20 21.60 17.07 20.80 21.60 ST (qr) 26.93 24.27 26.13 24.53 26.67 26.93 23.47 25.87 20.00 24.00 24.88 Boundary-SFT 22.13 22.40 21.33 24.00 22.67 21.87 22.67 19.73 15.47 16.00 20.83 Q-Align 20.53 19.73 19.47 19.47 22.93 20.80 21.87 19.47 8.00 9.33 18.16 GRPO 32.53 29.60 24.27 32.53 30.40 29.87 28.27 27.47 14.67 14.13 26.37 TAPO 27.73 22.67 21.33 28.00 25.07 26.93 22.67 23.73 9.33 9.33 21.68 Luar 28.80 27.73 29.33 30.67 29.33 31.73 29.07 27.20 20.80 24.00 27.87 Table 13: Per-language task accuracy (%) on PolyMath-High. Numbers are mean accuracy over seeds. Cell colors indicate changes from the Base row, clipped at \pm 20 points. Method Seen Unseen Avg. ar th sw en zh ko bn te yo zu Qwen3-4B 72.11 71.73 50.71 75.10 71.34 72.37 70.95 67.32 42.28 40.99 63.49 Full-Translation 74.71 75.10 72.11 75.10 76.13 74.58 76.01 73.02 64.33 68.35 72.94 Self-Assessment 73.15 72.24 50.06 74.84 74.71 73.02 71.73 68.35 41.63 43.84 64.36 Native-Tool-Use 73.28 73.93 54.60 76.52 69.65 72.63 72.24 67.96 46.43 50.45 65.77 ST (q) 72.11 72.24 69.91 75.23 76.65 74.45 72.63 72.89 64.72 67.19 71.80 ST (qr) 73.93 74.84 71.47 75.75 75.36 73.54 72.37 71.73 61.74 64.98 71.57 Boundary-SFT 71.60 70.43 53.89 71.60 65.37 71.79 69.84 62.65 42.41 56.42 63.60 Q-Align 68.09 69.78 52.27 72.37 70.04 67.57 66.67 63.81 31.26 36.06 59.79 GRPO 68.48 68.74 46.95 73.41 72.37 70.56 69.13 65.63 38.78 40.08 61.41 TAPO 68.87 71.34 55.25 74.58 71.98 69.39 67.32 64.85 32.81 36.06 61.25 Luar 71.47 74.19 69.26 75.36 72.11 72.37 70.30 67.32 46.69 65.24 68.43 Qwen3-8B 76.52 76.91 56.81 76.01 75.62 74.58 73.93 69.00 42.80 44.62 66.68 Full-Translation 79.38 80.03 75.75 76.01 78.86 77.69 76.78 76.91 67.32 70.17 75.89 Self-Assessment 76.52 77.56 61.48 78.47 77.95 75.75 74.84 72.63 46.69 47.86 68.98 Native-Tool-Use 74.58 64.98 67.57 74.97 72.50 73.41 77.04 70.82 63.81 66.80 70.65 ST (q) 76.01 74.71 53.44 77.17 76.13 76.26 73.54 71.34 45.27 46.04 66.99 ST (qr) 78.86 78.86 74.06 80.16 78.73 77.56 77.95 74.45 65.63 67.32 75.36 Boundary-SFT 73.80 72.50 55.38 74.45 70.82 72.89 70.69 68.87 46.43 51.75 65.76 Q-Align 72.24 73.93 56.03 76.78 73.15 72.11 69.91 68.48 36.45 41.37 64.05 GRPO 76.26 76.39 55.38 78.99 75.88 75.23 74.58 72.24 43.19 44.62 67.28 TAPO 68.48 58.75 50.06 73.67 62.91 64.33 63.68 62.52 28.40 33.98 56.68 Luar 75.75 76.39 65.24 75.36 75.88 75.23 71.47 70.30 54.09 61.09 70.08 Table 14: Per-language task accuracy (%) on MMLU-ProX-Lite. Numbers are mean accuracy over seeds. Cell colors indicate changes from the Base row, clipped at \pm 20 points.