mlsanityData to 14 Oct 2026

What Does Post-Training Change in Multilingual Reasoning?

arXiv 2609.37104 ↗First posted 29 Sep 2026cs.CL cs.AIcited by 0

On the map: LLM reasoning → Reasoning benchmarks and evaluation

Abstract (arXiv)

Open-source reasoning models provide unequal access to reasoning capability across languages. When a model can solve a problem but cannot deliver a complete solution in the user's language, language becomes an access barrier rather than merely a source of performance variation. We audit Qwen3 checkpoints on competition-mathematics tasks in eleven languages. Across the ten non-English languages, only 15.4-17.9% of problems receive a correct, terminating solution with visible reasoning in the requested language in any of 16 samples, compared with 92.9% in English. To identify the source of this disparity, we evaluate thirteen endpoints from one model family, spanning released checkpoints, multilingual supervised fine-tuning (SFT) at two scales, controlled SFT ablations, and three reinforcement-learning (RL) reward formulations. We jointly track correctness, language adherence, termination, and delivery efficiency. The dominant bottleneck shifts across post-training stages. Released models often reason in English. Multilingual SFT restores target-language reasoning, but accuracy declines across multilingual, English-only, and single-language SFT runs, showing that this cost is not specific to multilingual mixing; non-English reasoning traces additionally become prone to non-terminating loops. RL restores termination in both arms at no cost in accuracy, but only the arm whose reward includes a language term delivers: rewarding correctness alone returns the model to English. Together, these stages establish a constructive post-training path from English-pivoted capability to multilingual reasoning that is reliably delivered.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning. Builds on 19 of its 23 references that are in the map. First posted on arXiv 29 Sep 2026.

Builds on

19 of 23 references.

Closest ancestors · 8

Same direction, newest first

  1. 2026LANG: Reinforcement Learning for Multilingual Reasoning with Language-Adaptive Hint Guidance21 May
  2. 2026ReasonXL: Shifting LLM Reasoning Language Without Sacrificing Performance14 Apr
  3. 2025Think Natively: Unlocking Multilingual Reasoning with Consistency-Enhanced Reinforcement Learning8 Oct
  4. 2025The Impact of Language Mixing on Bilingual LLM Reasoning21 Jul
  5. 2025Learn Globally, Speak Locally: Bridging the Gaps in Multilingual Reasoning7 Jul
  6. 2025When Models Reason in Your Language: Controlling Thinking Language Comes at the Cost of Accuracy28 May
  7. 2025Qwen3 Technical Report14 May
  8. 2025Do Multilingual LLMs Think In English?21 Feb

Other references · 11

  1. 2017Proximal Policy Optimization Algorithmscited by 15,215
  2. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scalecited by 2,810
  3. 2015High-Dimensional Continuous Control Using Generalized Advantage Estimationcited by 2,026
  4. 2024Do Llamas Work in English? On the Latent Language of Multilingual Transformerscited by 281
  5. 2023Language Model Tokenizers Introduce Unfairness Between Languagescited by 214
  6. 2025PolyMath: Evaluating Mathematical Reasoning in Multilingual Contextscited by 56
  7. 2025Wait, Wait, Wait... Why Do Reasoning Models Loop?cited by 26
  8. 2025Long Chain-of-Thought Reasoning Across Languagescited by 19
  9. 2025The Token Tax: Systematic Bias in Multilingual Tokenizationcited by 14
  10. 2025CoCo-CoLa: Evaluating and Improving Language Adherence in Multilingual LLMscited by 4
  11. 2026Disentangling Language Roles in Multilingual LLM Task Executioncited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Rethinking the Multilingual Reasoning Gap with Layer Swap
  2. 2025Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models?
  3. 2025Beyond English-Centric Training: How Reinforcement Learning Improves Cross-Lingual Reasoning in LLMs
  4. 2025Parallel Scaling Law: Unveiling Reasoning Generalization through A Cross-Linguistic Perspective
  5. 2025Best-of-L: Cross-Lingual Reward Modeling for Mathematical Reasoning
  6. 2025Demystifying Multilingual Chain-of-Thought in Process Reward Modeling
  7. 2025Cross-lingual Collapse: How Language-Centric Foundation Models Shape Reasoning in Large Language Models
  8. 2026What Makes Good Multilingual Reasoning? Disentangling Traces with Measurable Features