Computation and Language · arXiv:2608.13698 v1 · submitted 13 Aug 2026 · 71 references
GRPO Beyond English: A Large-Scale Study of GRPO in Non-English and Multilingual Settings
The paper says
“Reinforcement Learning with Verifiable Rewards (RLVR), often optimized with Group Relative Policy Optimization (GRPO), has become a central recipe for improving the reasoning capabilities of pretrained language models but current studies remain heavily English-centric.”
more of the abstract
We conduct a large-scale empirical study of multilingual and non-English GRPO across a wide range of base models, training languages, and different reasoning language rewards. We find that training to reason in the native language often leaves only a small gap to training for English reasoning. We further observe strong crosslingual transfer: training in one language often improves performance in many others. However, specific trends are highly model- and language-dependent. In some cases, training in a particular language induces severe regressions on out-of-domain capabilities in other languages. Our analysis shows that RLVR beyond English can provide broad crosslingual gains, but also requires broad evaluation to detect language-specific regressions.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 27 Aug 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. A genuinely broad empirical sweep of GRPO in 11 languages, but the transfer and native-reasoning-gap claims are anchored to template tasks and do not survive on hard natural-language math; the regression warnings are the real payload. This line is wrong
3 major objections · 3 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 71 references say.
The paper says, at Abstract; §4.2; Table 2
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The central claim of strong crosslingual transfer and broad RLVR gains is demonstrated mainly on the Multilingual Reasoning Gym, whose training and evaluation tasks are template/procedural and deliberately exclude English-content word problems. Table 2 shows that on natural-language math benchmarks the pattern is much weaker or reversed: for Qwen3-1.7B/4B/8B, English-reward training loses 12.0–15.7 pp on Hard Math in the training language and 14.3–20.7 pp in crosslingual evaluation, despite large gains on MRG seen tasks in Table 1.
more of objection 1
The abstract and conclusion state that 'RLVR beyond English can provide broad crosslingual gains' without restricting this claim to template-based tasks. Please either scope the headline claims to the task distribution on which they were established or provide separate crosslingual-transfer analyses for natural-language math tasks.
The paper says, at Section 3; Limitations; Figures 2–3
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
All configurations use one shared recipe (500 steps, learning rate 1e-6, batch 512, no KL penalty) and a single seed per configuration. The quantitative claims such as 'English-reasoning rewards are 1.3–4.1 pp better than native-reasoning rewards' and 'strong crosslingual transfer' therefore have no variance estimates and may depend on the particular recipe and seed.
more of objection 2
The Limitations section itself acknowledges that 'some negative results... may therefore be reduced by targeted hyperparameter tuning' and that 'we run all experiments with a single seed only.' For the central claims to be load-bearing, the authors should add at least a small multi-seed and/or multi-recipe sensitivity study for representative model-language pairs, or explicitly reframe the conclusions as descriptive of this single-recipe setup rather than as properties of GRPO in general.
The paper says, at Section 4.1; Appendix B
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The English-versus-native reward comparison is restricted to model-language pairs in which both rewards successfully control the rollout language. Appendix B provides a welcome sensitivity analysis showing the aggregate advantage decreases only from 2.8 to 2.4 pp on seen languages when all pairs are included, so this selection is not the main problem.
more of objection 3
I would still ask the authors to report, alongside Figures 2 and 3, the number of pairs contributing to each bar and the per-pair spread, because many of the stated conclusions are averages over heterogeneous model-language combinations that can differ by more than 10 pp in individual cases.
3 minor objections
The paper says, at Section 4.2; Table 2
Pith says: minor objection 1 of 3
The definitions of 'Seen language' and 'Crosslingual Transfer' in Table 2 should be stated in the table caption or the main text; the current caption reports averages over monolingual runs but does not specify whether 'Seen' averages only the training language or includes the multilingual condition.
The paper says, at Figure 1
Pith says: minor objection 2 of 3
Figure 1 would be easier to interpret if the caption stated the number of rollouts per bucket and the set of languages used for the language detector, since the distinction between 'prompt language' and 'other' depends on the detector's language inventory.
show 1 more
The paper says, at General
Pith says: minor objection 3 of 3
The text contains many encoding and typographical artifacts (e.g., 'difficult', 'A vg@8', 'T rain') that should be cleaned up in a final proofreading pass.
What would settle it
Compute the Table 1 and Figure 2 aggregations separately for (a) MRG template tasks, (b) Easy Math (MGSM + PolyMath-low), and (c) Hard Math (PolyMath-medium/high/top), and report them per model family (Base vs instruction-tuned). If the mean cross-lingual delta on (c) is negative for the non-Base models, as Table 2 already suggests, the conclusion should be restated as 'broad transfer on template/procedural tasks and easy math, with regression on hard natural-language math.' A stronger decisive check: run the same 500-step English-reward recipe on a non-template multilingual reasoning training set, e.g., translated MGSM, for 2–3 models, and compare the cross-lingual transfer and the native-vs-English reward gap against the MRG-trained checkpoints; if transfer drops materially, the MRG-based findings are task-specific.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 27 Aug 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 3 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
This is a genuinely useful breadth-first measurement of GRPO outside English: nine models, eleven training languages, English- and native-reasoning rewards, monolingual and multilingual runs. The paper's best result is the language-dependent regression: specific model-language pairs lose large amounts on unseen English-content tasks or hard math even when they gain in the training language. That pattern is consistent, appears on both the reasoning-gym tasks and MGSM/PolyMath, and is the finding I would trust most. The authors also deserve credit for the Appendix B sensitivity analysis and for stating the single-seed and shared-recipe limitations instead of burying them.
The headline claims need more care than the abstract gives them. The 'small gap' between native and English reasoning rewards and the 'strong crosslingual transfer' are mostly measured on the Multilingual Reasoning Gym tasks in Table 1, which are procedural/template tasks: arithmetic, sorting, mazes, graph coloring. On those tasks language is a thin surface, so transfer is almost free. Once you move to natural-language math in MGSM and PolyMath, Table 2 tells a different story: the non-Base Qwen3 models, which show big gains on the gym tasks, are flat on Easy Math and lose 12-21 points on Hard Math under English rewards. The stress-test note is right: the broad 'RLVR beyond English works' statement is task-specific, and the paper should either scope the conclusion to procedural tasks or show that transfer holds on harder, language-bound benchmarks.
The soft spots are real but proportionate. One seed, one shared hyperparameter recipe, no KL: that makes the exact deltas provisional. No code or data release that I can see, which is a problem for an empirical matrix this size. The Qwen3-14B task filter could bias the task set toward what one family can do. None of this sinks the regression findings, but it should keep the quantitative claims modest.
I would send this to a serious referee. A revised version with narrowed claims, released code and data, and at least a handful of multi-seed or variance checks would be a useful citable reference. The citation pattern seems fair, including the self-citation to their own reasoning gym, which is legitimate since it is the training environment.
The strongest objection
The central claim is that GRPO trained in one language transfers broadly, and that native-reasoning rewards cost little relative to English-reasoning rewards. The primary evidence for both is Table 1 and Figures 2–3, computed on the Multilingual Reasoning Gym: the 62 training tasks and 30 unseen tasks are template/procedural tasks (Appendix A: basic_arithmetic, bitwise_arithmetic, number_sorting, maze, sudoku, course_schedule, etc.), and English-content word problems are deliberately excluded from training. On such tasks language is a shallow surface: a procedure learned in one language transfers to another almost for free, and the language of the reasoning trace has weak coupling to the actual problem-solving content. The natural-language benchmarks included for evaluation, MGSM and PolyMath, tell a different story. In Table 2, the non-Base Qwen3-1.7B/4B/8B models lose 12.0–15.7 pp on Hard Math in the training language and 14.3–20.7 pp in cross-lingual evaluation under English rewards, while their Easy Math gains are near zero; the same models show large MRG gains in Table 1. The positive 'strong crosslingual transfer' claim is therefore not demonstrated as a property of GRPO or of language in general; it is observed on a specific, largely symbolic task distribution, mostly for Base models. Since the abstract and conclusion state the broader result without this scope restriction, the headline is at risk of task-specific overgeneralization. This is a construct-validity issue rather than an internal inconsistency: within the MRG setting the controlled comparisons seem sound, but the inference from MRG to 'RLVR beyond English' is the weakest link.
I read the paper as a controlled empirical mapping rather than a mechanistic theory of multilinguality. Its internal comparisons are carefully executed: reasoning-language control is measured, Appendix B provides a sensitivity analysis of the language-control filtering, three languages are held out, and the limitations section honestly flags the single-seed and shared-recipe issues. I do not think the paper is wrong about what happens on the Multilingual Reasoning Gym; I think the weakest step is the inference from that task distribution to the broad claim that GRPO offers strong crosslingual transfer and only a small native-reasoning penalty. The reader's single-seed/hyperparameter concern is real but, in my view, less decisive for the central claims: many reported effects are large (20–30 pp transfer gains, 10–20 pp regressions) and recur across nine models, so they are unlikely to be pure seed noise; the single-seed issue mainly threatens the small 1–4 pp gap estimates, not the existence of transfer. By contrast, the template-based nature of the training tasks affects the construct validity of both positive claims, because transfer on symbolic tasks may not transfer to linguistically rich reasoning. Tables 2 and 3 already show markedly weaker or negative outcomes on MGSM/PolyMath and English-content tasks for several models. Thus the CONDITIONAL verdict is appropriate: the paper should qualify its conclusions to the evaluated task distribution and provide the restricted aggregations and the non-template training check described above before the headline is accepted as a general property of GRPO. I keep the reader's verdict unchanged, since my concern supports the existing conditional rather than moving it to accept or reject.
The referee's report
If the findings hold, they provide a practical and comparatively optimistic answer to whether RLVR can be applied to non-English reasoning: native-language training would be nearly as effective as English-reasoning training, and a single training language would transfer broadly, provided evaluation covers many languages. The study's strength is its breadth: nine model families, eleven training languages, heldout languages and tasks, natural-language math benchmarks, and a sensitivity analysis in Appendix B. The paper is an empirical measurement study, not a derivation, and its results are anchored to external benchmarks and held-out tasks. However, the headline claims are currently broader than the evidence, because the transfer results are dominated by template-based procedural tasks and all configurations share a single hyperparameter recipe and a single seed.
Pith's reading of the paper (a machine's summary, not the authors')
This paper asks whether reinforcement learning with verifiable rewards (RLVR), in the specific form of Group Relative Policy Optimization, can be moved out of its English-centric comfort zone without a large accuracy penalty. Across nine base models and eleven training languages, it finds that rewarding a model for reasoning in the prompt's native language costs only a modest average gap — roughly 1.3 to 4.1 percentage points — compared with rewarding English reasoning, far below the double-digit penalties reported when native reasoning is imposed without sustained training. It also documents strong crosslingual transfer: training in a single language typically lifts performance in many other languages, including languages never seen during training, and English is often not the best transfer source. The caveat is that these averages hide severe, model- and language-specific regressions — for example, Qwen3 (non-Base) models trained on Swahili or Telugu collapse on unseen English-content tasks, and difficult-math performance drops across many models when English is absent from the training mix. The paper's conclusion is that non-English RLVR can deliver broad gains, but only if evaluation covers many languages and task types, since transfer numbers alone will not reveal the regressions.
What carries the argument. The load-bearing mechanism is reward shaping inside Group Relative Policy Optimization (GRPO), the reinforcement-learning algorithm that updates a policy from group-relative advantages on automatically verifiable outcomes. The study holds the GRPO training recipe fixed and varies two things: the language of the prompts (separate monolingual runs in eleven languages plus one multilingual run) and a reasoning-language reward that pays a bonus when the model's reasoning traces match either the prompt language ('native') or English. Because the language reward is paid only when the extracted answer is correct and weighted at half the correctness reward, it separates the decision of what language to reason in from whether the answer is right. The complementary piece is the Multilingual Reasoning Gym, a procedurally generated task environment with a dynamic difficulty curriculum (difficulty rises past 70 percent correct on the last 32 rollouts and falls below 10 percent), which keeps task difficulty calibrated per language and supplies the verifiable math, logic, and puzzle tasks used for training and evaluation. The design also evaluates every checkpoint in every language on both seen and held-out tasks, which is what makes the language-specific regressions visible rather than averaged away.
Core claim. Under a fixed GRPO recipe (a DAPO-style policy loss with no KL penalty, 500 steps), the paper establishes three results. First, a reasoning-language reward can reliably steer the language of model reasoning: English rewards produce English traces in roughly 90 to 99 percent of rollouts, and native rewards produce prompt-language traces in about 95 percent of rollouts on seen languages, without sustained accuracy collapse. Second, native-language rewards trail English rewards by only 1.3 to 4.1 percentage points on average across models on seen languages, with the largest gaps on lower-resource languages such as Bengali, Telugu, and Swahili. Third, crosslingual transfer under GRPO is strong: same-language training usually remains best in-domain, but much of the gain transfers to other languages, training in Bengali can be the best transfer source into Chinese for some Qwen3 models, and multilingual training nearly matches an oracle that picks the best transfer language per target for Qwen3-Base. The same experiments reveal a failure mode: for Qwen3 (non-Base) models, training on Swahili or Telugu induces double-digit regressions on unseen English-content tasks even while the training language improves, and Hard Math performance falls across many models whenever English prompts are absent from training. The paper reads these results as evidence that RLVR beyond English can provide broad crosslingual gains, but that broad evaluation is required to detect language-specific regressions.
The premise it rests on. If the transfer gains and language-specific regressions reported here are peculiar to the study's one shared recipe — 500 steps, learning rate 1e-6, batch 512, no KL penalty, and a single seed per configuration — rather than properties of GRPO and the training language themselves, the central claims would not generalize; the authors concede as much in their Limitations, noting that 'some negative results, including cases of language-dependent task regressions, may therefore be reduced by targeted hyperparameter tuning.'
If the paper is right (Pith's inferences, not the authors')
Non-English RLVR becomes practical: teams that need models to reason in a target language can train in that language directly and expect an average cost of only a few percentage points over English-reasoning training, instead of the double-digit penalty suggested by earlier small-scale work.
English is not the default best transfer source: because training in Bengali, French, or German can transfer to some targets better than English can, practitioners should select training languages using measured transfer performance rather than defaulting to English.
Multilingual training can replace a set of per-language models: for Qwen3-Base models it lands within 0.3 percentage points of an oracle that chooses the best transfer language separately for each target language, at the cost of a single checkpoint.
Evaluation must be broad and disaggregated: the same checkpoint that gains tens of points in its training language can lose double-digit points on unseen English-content tasks or on hard math, so average gains across languages are an insufficient success metric.
Training-mix design is a control lever: including English prompts in the training data, as the English-monolingual and multilingual runs do, largely prevents the Hard Math regression, which gives a concrete rule for composing multilingual training data.
If the regressions are partly an artifact of the fixed recipe rather than intrinsic to GRPO plus language — a possibility the authors themselves leave open — then adding a mild KL penalty or per-language learning rates could rehabilitate Swahili and Telugu as training languages; this is a cheap experiment the paper did not run.
The transfer pattern (Bengali to Chinese being best for some Qwen3 models, Telugu and Swahili transferring poorly) suggests transfer is mediated by the geometry of shared representation space in the base model rather than by language family or script; a testable extension would try to predict the full training-by-evaluation transfer matrix from a base model's inter-language similarity.
A practical recipe that keeps a small fixed fraction of English prompts inside every non-English training run could capture most of the crosslingual gains while avoiding the Hard Math and English-content regressions, trading away the clean monolingual separation the paper deliberately preserved.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Reinforcement Learning with Verifiable Rewards (RLVR), often optimized with Group Relative Policy Optimization (GRPO), has become a central recipe for improving the reasoning capabilities of pretrained language models but current studies remain heavily English-centric. We conduct a large-scale empirical study of multilingual and non-English GRPO across a wide range of base models, training languages, and different reasoning language rewards. We find that training to reason in the native language often leaves only a small gap to training for English reasoning. We further observe strong crosslingual transfer: training in one language often improves performance in many others. However, specific trends are highly model- and language-dependent. In some cases, training in a particular language induces severe regressions on out-of-domain capabilities in other languages. Our analysis shows that RLVR beyond English can provide broad crosslingual gains, but also requires broad evaluation to detect language-specific regressions.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23
The authors' main claim, as the classifier picked it out 1.00
“Our analysis shows that RLVR beyond English can provide broad crosslingual gains, but also requires broad evaluation to detect language-specific regressions.”
How the main claim is stated: asserted 0.95
What the abstract itself states, as probabilities: claims first 0.08 · states limitation 0.86 · human participants 0.04 · controlled comparison 0.48 · randomized assignment 0.02 · states quantitative result 0.06 · compares with prior results 0.56 · mentions code or data availability 0.02
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Reinforcement Learning with Verifiable Rewards (RLVR), often optimized with Group Relative Policy Optimization (GRPO), has become a central recipe for improving the reasoning capabilities of pretrained language models but current studies remain heavily English-centric. | background 0.75 | 0.05 | 0.04 | asserted 1.00 |
| We conduct a large-scale empirical study of multilingual and non-English GRPO across a wide range of base models, training languages, and different reasoning language rewards. | contribution 0.95 | 0.96 | 0.04 | asserted 0.95 |
| We find that training to reason in the native language often leaves only a small gap to training for English reasoning. | result 1.00 | 0.96 | 0.19 | asserted 1.00 |
| We further observe strong crosslingual transfer: training in one language often improves performance in many others. | result 1.00 | 0.94 | 0.14 | asserted 0.94 |
| However, specific trends are highly model- and language-dependent. | result 0.95 | 0.86 | 0.07 | asserted 0.97 |
| In some cases, training in a particular language induces severe regressions on out-of-domain capabilities in other languages. | result 0.99 | 0.88 | 0.08 | asserted 0.68 |
| Our analysis shows that RLVR beyond English can provide broad crosslingual gains, but also requires broad evaluation to detect language-specific regressions. | interpretation 0.98 | 0.94 | 0.87 | asserted 0.95 |
What kind of paper the classifier reports this is: empirical study 1.00 · evidence from observation or experiment 1.00
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 5,532 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
References (71)
- [1] Proximal Policy Optimization Algorithms , author =. 2017 , eprint = on Pith
- [2] Accelerate: Training and inference at scale made simple, efficient and adaptable. , author = on Pith
- [3] Efficient Memory Management for Large Language Model Serving with PagedAttention , author =. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year = on Pith
- [4] Reinforcement Fine-Tuning Naturally Mitigates Forgetting in Continual Post-Training , author =. 2026 , eprint = on Pith
- [5] Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting , author =. 2025 , eprint = on Pith
- [6] Adam: A Method for Stochastic Optimization , author =. 2017 , eprint = on Pith
- [7] Decoupled Weight Decay Regularization , author =. 2019 , eprint = on Pith
- [8] Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models? , author =. 2025 , eprint = on Pith
- [10] Rust, Phillip and Pfeiffer, Jonas and Vulić, Ivan and Ruder, Sebastian and Gurevych, Iryna , editor =. How. Proceedings of the 59th. 2021 , pages =. doi:10.18653/v1/2021.acl-long.243 , abstract = on Pithdoi:10.18653/v1/2021.acl-long.243
- [11] Ahia, Orevaoghene and Kumar, Sachin and Gonen, Hila and Kasai, Jungo and Mortensen, David R. and Smith, Noah A. and Tsvetkov, Yulia , month = may, year =. Do. doi:10.48550/arXiv.2305.13707 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2305.13707
- [12] Yamaguchi, Atsuki and Villavicencio, Aline and Aletras, Nikolaos , month = sep, year =. An. doi:10.48550/arXiv.2402.10712 , abstract = on PithPith's reviewarXivdoi:10.48550/arxiv.2402.10712
- [13] Ali, Mehdi and Fromm, Michael and Thellmann, Klaudia and Rutmann, Richard and Lübbering, Max and Leveling, Johannes and Klug, Katrin and Ebert, Jan and Doll, Niclas and Buschhoff, Jasper and Jain, Charvi and Weber, Alexander and Jurkschat, Lena and Abdelwahab, Hammam and John,... on Pithdoi:10.18653/v1/2024.findin
- [14] RL's Razor: Why Online Reinforcement Learning Forgets Less , author =. 2025 , eprint = on Pith
- [15] Spurious Rewards: Rethinking Training Signals in RLVR , author =. 2026 , eprint = on Pith
- [16] Understanding R1-Zero-Like Training: A Critical Perspective , author =. 2025 , eprint = on Pith
- [17] DAPO: An Open-Source LLM Reinforcement Learning System at Scale , author =. 2025 , eprint = on Pith
- [18] Crosslingual Reasoning through Test-Time Scaling , author =. 2025 , eprint = on Pith
- [19] Do Llamas Work in English? On the Latent Language of Multilingual Transformers , author =. 2024 , eprint = on Pith
- [20] Linguistic Generalizability of Test-Time Scaling in Mathematical Reasoning , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month = jul, year =. doi:10.18653/v1/2025.acl-long.699 , pages = on Pithdoi:10.18653/v1/2025.acl-long.699
- [21] Multilingual Reasoning Gym: Multilingual Scaling of Procedural Reasoning Environments , author=. 2026 , eprint= on Pith
show 51 more
- [22] Shulin Huang and Yiran Ding and Junshu Pan and Yue Zhang , booktitle =. Beyond English-Centric Training: How Reinforcement Learning Improves Cross-Lingual Reasoning in. 2026 , url = on Pith
- [23] Breaking Language Barriers in Multilingual Mathematical Reasoning: Insights and Observations , author =. 2023 , eprint = on Pith
- [24] Learn Globally, Speak Locally: Bridging the Gaps in Multilingual Reasoning , author =. 2025 , eprint = on Pith
- [26] Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Zhang, Ruoyu and Ma, Shirong and Bi, Xiao and Zhang, Xiaokang and Yu, Xingkai and Wu, Yu and Wu, Z. F. and Gou, Zhibin and Shao, Zhihong and Li, Zhuoshu and Gao, Zi... on Pith
- [27] Cross-lingual Collapse: How Language-Centric Foundation Models Shape Reasoning in Large Language Models , author =. 2025 , eprint = on Pith
- [28] DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author =. arXiv preprint arXiv:2402.03300 , year = on PithPith's reviewarXiv
- [29] Stojanovski, Zafir and Stanley, Oliver and Sharratt, Joe and Jones, Richard and Adefioye, Abdulhakeem and Kaddour, Jean and K. arXiv preprint arXiv:2505.24760 , year = on PitharXiv
- [30] When Verifiable Rewards Switch the Language: Cross-lingual Collapse in Chain-of-Thought , author =. arXiv preprint arXiv:2506.05850 , year = on PitharXiv
- [31] When Models Reason in Your Language: Controlling Thinking Language Comes at the Cost of Accuracy , author =. Findings of the Association for Computational Linguistics: EMNLP 2025 , year =. doi:10.18653/v1/2025.findings-emnlp.1103 , url = on Pithdoi:10.18653/v1/2025.findings-emnlp.1103
- [32] Language Mixing in Reasoning Language Models: Patterns, Impact, and Internal Causes , author =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , year =. doi:10.18653/v1/2025.emnlp-main.132 , url = on Pithdoi:10.18653/v1/2025.emnlp-main.132
- [33] Crosslingual Reasoning through Test-Time Scaling , author =. arXiv preprint arXiv:2505.05408 , year = on PithPith's reviewarXiv
- [34] Think Natively: Unlocking Multilingual Reasoning with Consistency-Enhanced Reinforcement Learning , author =. arXiv preprint arXiv:2510.07300 , year = on PitharXiv
- [35] Aligning Multilingual Reasoning with Verifiable Semantics from a High-Resource Expert Model , author =. arXiv preprint arXiv:2509.25543 , year = on PitharXiv
- [36] Gained In Translation: Privileged Pairwise Judges Enhance Multilingual Reasoning , author =. arXiv preprint arXiv:2601.18722 , year = on PitharXiv
- [37] mGRPO: Unlocking LLM Reasoning through Multilingual Thinking , author =. International Conference on Learning Representations (ICLR) 2026 (submission) , year = on Pith
- [38] ExpLang: Improved Exploration and Exploitation in LLM Reasoning with On-Policy Thinking Language Selection , author =. arXiv preprint arXiv:2602.21887 , year = on PitharXiv
- [39] An, Chenxin and Xie, Zhihui and Li, Xiaonan and Li, Lei and Zhang, Jun and Gong, Shansan and Zhong, Ming and Xu, Jingjing and Qiu, Xipeng and Wang, Mingxuan and Kong, Lingpeng , howpublished =. 2025 , url = on Pith
- [40] Zhang, Jixiao and Zuo, Chunsheng , booktitle =. 2025 , month = nov, address =. doi:10.18653/v1/2025.emnlp-main.287 , url = on Pithdoi:10.18653/v1/2025.emnlp-main.287
- [41] xCoT: Cross-lingual Instruction Tuning for Cross-lingual Chain-of-Thought Reasoning , author =. arXiv preprint arXiv:2401.07037 , year = on PithPith's reviewarXiv
- [42] Language Models are Multilingual Chain-of-Thought Reasoners , author =. arXiv preprint arXiv:2210.03057 , year = on PithPith's reviewarXiv
- [43] Breaking Language Barriers in Multilingual Mathematical Reasoning: Insights and Observations , author =. Findings of the Association for Computational Linguistics: EMNLP 2024 , year =. doi:10.18653/v1/2024.findings-emnlp.411 , url = on Pithdoi:10.18653/v1/2024.findings-emnlp.411
- [44] Yoon, Dongkeun and Jang, Joel and Kim, Sungdong and Kim, Seungone and Shafayat, Sheikh and Seo, Minjoon , booktitle =. 2024 , month = aug, address =. doi:10.18653/v1/2024.acl-long.405 , url = on Pithdoi:10.18653/v1/2024.acl-long.405
- [45] Multilingual Reasoning via Self-training , author =. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , year =. doi:10.18653/v1/2025.naacl-long.577 , url = on Pithdoi:10.18653/v1/2025.naacl-long.577
- [46] Luo, Wenyang and Zhao, Wayne Xin and Sha, Jing and Wang, Shijin and Wen, Ji-Rong , booktitle =. 2025 , month = nov, address =. doi:10.18653/v1/2025.findings-emnlp.598 , url = on Pithdoi:10.18653/v1/2025.findings-emnlp.598
- [47] Wang, Yiming and Zhang, Pei and Tang, Jialong and Wei, Hao-Ran and Yang, Baosong and Wang, Rui and Sun, Chenshu and Sun, Feitong and Zhang, Jiran and Wu, Junxuan and Cang, Qiqian and Zhang, Yichang and Lin, Junyang and Huang, Fei and Zhou, Jingren , booktitle =. PolyMath: Eval... on Pith
- [48] She, Shuaijie and Zou, Wei and Huang, Shujian and Zhu, Wenhao and Liu, Xiang and Geng, Xiang and Chen, Jiajun , booktitle =. 2024 , month = aug, address =. doi:10.18653/v1/2024.acl-long.539 , url = on Pithdoi:10.18653/v1/2024.acl-long.539
- [49] Learn Globally, Speak Locally: Bridging the Gaps in Multilingual Reasoning , author =. arXiv preprint arXiv:2507.05418 , year = on PitharXiv
- [50] Cross-lingual Prompting: Improving Zero-shot Chain-of-Thought Reasoning across Languages , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year =. doi:10.18653/v1/2023.emnlp-main.163 , url = on Pithdoi:10.18653/v1/2023.emnlp-main.163
- [51] SLAM: Towards Efficient Multilingual Reasoning via Selective Language Alignment , author =. 2025 , eprint = on Pith
- [52] mCoT: Multilingual Instruction Tuning for Reasoning Consistency in Language Models , author =. 2024 , eprint = on Pith
- [53] Overcoming Catastrophic Forgetting in Zero-Shot Cross-Lingual Generation , author =. 2022 , eprint = on Pith
- [54] Measuring Catastrophic Forgetting in Cross-Lingual Transfer Paradigms: Exploring Tuning Strategies , author =. 2025 , eprint = on Pith
- [55] Conditions for Catastrophic Forgetting in Multilingual Translation , author=. 2025 , eprint= on Pith
- [56] MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention , author=. 2025 , eprint= on Pith
- [57] Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model , author=. 2025 , eprint= on Pith
- [58] Skywork Open Reasoner 1 Technical Report , author=. 2025 , eprint= on Pith
- [59] Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning , author=. 2026 , eprint= on Pith
- [60] Saji, Alan and Dabre, Raj and Kunchukuttan, Anoop and Puduppully, Ratish. The Reasoning Lingua Franca: A Double-Edged Sword for Multilingual AI. Proceedings of the 19th Conference of the E uropean Chapter of the A ssociation for C omputational L inguistics (Volume 2: Short Pap... on Pithdoi:10.18653/v1/2026.eacl-short.25
- [61] Language as a Latent Variable for Reasoning Optimization , author =. 2026 , eprint =. doi:10.48550/arXiv.2604.21593 , url = on PithPith's reviewarXivdoi:10.48550/arxiv.2604.21593
- [62] Huang, Xu and Lai, Zhejian and Huang, Zixian and Chen, Jiajun and Huang, Shujian , year =. doi:10.48550/arXiv.2603.25419 , url =. 2603.25419 , archivePrefix = on Pithdoi:10.48550/arxiv.2603.25419
- [63] Self-Improving Multilingual Long Reasoning via Translation-Reasoning Integrated Training , author =. 2026 , eprint =. doi:10.48550/arXiv.2602.05940 , url = on PithPith's reviewarXivdoi:10.48550/arxiv.2602.05940
- [64] Crosslingual On-Policy Self-Distillation for Multilingual Reasoning , author =. 2026 , eprint =. doi:10.48550/arXiv.2605.09548 , url = on PithPith's reviewarXivdoi:10.48550/arxiv.2605.09548
- [65] Cross-lingual Self-Consistency for Multilingual Reasoning with Language Models , author =. 2026 , eprint =. doi:10.48550/arXiv.2606.01464 , url = on PithPith's reviewarXivdoi:10.48550/arxiv.2606.01464
- [66] Improving Cross-Lingual Factual Recall via Consistency-Driven Reinforcement Learning , author =. 2026 , eprint =. doi:10.48550/arXiv.2606.06586 , url = on PithPith's reviewarXivdoi:10.48550/arxiv.2606.06586
- [67] Language Detection Library for Java , author = on Pith
- [68] Kydlíček, Hynek , title =. 2025 , howpublished = on Pith
- [69] Tulu 3: Pushing Frontiers in Open Language Model Post-Training , author=. 2025 , eprint= on Pith
- [70] Kimi k1.5: Scaling Reinforcement Learning with LLMs , author=. 2025 , eprint= on Pith
- [71] Qwen3 Technical Report , author=. 2025 , eprint= on Pith
- [72] Bakouch, Elie and Ben Allal, Loubna and Lozhkov, Anton and Tazi, Nouamane and Tunstall, Lewis and Patiño, Carlos Miguel and Beeching, Edward and Roucher, Aymeric and Reedi, Aksel Joonas and Gallouédec, Quentin and Rasul, Kashif and Habib, Nathan and Fourrier, Clémentine and Ky... on Pith
- [73] Gemma 3 Technical Report , author=. 2025 , eprint= on Pith
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
- A Survey on RLOps for Large Language Models cs.SE · Nov 2026
How the literature uses it
What this paper does with the arXiv papers it cites
What it cites · 20 sentences without a confident label
Uses the method or tool of 4 papers
-
Our study uses the Multilingual Reasoning Gym [ref] for training, which provides procedurally generated, template-based RLVR tasks with a dynamic difficulty curriculum.
2603.10793 · Limitations · 0.87
-
Following standard RLVR practice, we give a binary correctness reward when the extracted answer is verified as correct [ref] .
Skywork Open Reasoner 1 Technical Report · Experimental Setup · 0.81
-
Following recent large-scale reasoning-RL recipes, we do not add an explicit KL penalty to the GRPO objective [ref] .
Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model · Language-Dependent Task Regressions · 0.86
Cites as support 5 papers
-
Recent evidence suggests that, under RLVR, reasoning traces may drift toward a dominant pre-training language, often English, even when prompts are posed in other languages [ref] .
2506.05850 · Introduction · 0.98
-
Recent studies further indicate that RL-based post-training can transfer improvements across languages, including languages excluded from training, and in some settings generalizes more effectively than supervised fine-tuning [ref] .
2507.05418 · Related Work · 0.98
-
Recent studies further indicate that RL-based post-training can transfer improvements across languages, including languages excluded from training, and in some settings generalizes more effectively than supervised fine-tuning [ref] .
2510.07300 · Related Work · 0.98
Cites as background 6 papers
-
Existing approaches differ operationally—through various methods such as translations, reference traces, judges, or teacher distributions—but all use different forms of crosslingual supervision signals to bridge the gap between English and non-English reasoning [ref] .
2509.25543 · Related Work · 0.85
-
Existing approaches differ operationally—through various methods such as translations, reference traces, judges, or teacher distributions—but all use different forms of crosslingual supervision signals to bridge the gap between English and non-English reasoning [ref] .
2510.07300 · Related Work · 0.85
-
Existing approaches differ operationally—through various methods such as translations, reference traces, judges, or teacher distributions—but all use different forms of crosslingual supervision signals to bridge the gap between English and non-English reasoning [ref] .
2601.18722 · Related Work · 0.85
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment