Cross-lingual Collapse: How Language-Centric Foundation Models Shape Reasoning in Large Language Models

• URL: http://arxiv.org/abs/2506.05850v2
• Date: Mon, 09 Jun 2025 11:55:27 GMT
• Title: Cross-lingual Collapse: How Language-Centric Foundation Models Shape Reasoning in Large Language Models
• Authors: Cheonbok Park, Jeonghoon Kim, Joosung Lee, Sanghwan Bae, Jaegul Choo, Kang Min Yoo,
• Abstract summary: We identify textbfCross-lingual Collapse, a systematic drift in which a multilingual language model reverts to its dominant pre-training language.<n>Our experiments reveal three key findings: (i) GRPO rapidly amplifies pre-training language imbalances, leading to the erosion of low-resource languages within just a few hundred updates; (ii) language consistency reward mitigates this drift but does so at the expense of an almost 5 - 10 pp drop in accuracy.
• Score: 44.94287386776289
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: We identify \textbf{Cross-lingual Collapse}, a systematic drift in which the chain-of-thought (CoT) of a multilingual language model reverts to its dominant pre-training language even when the prompt is expressed in a different language. Recent large language models (LLMs) with reinforcement learning with verifiable reward (RLVR) have achieved strong logical reasoning performances by exposing their intermediate reasoning traces, giving rise to large reasoning models (LRMs). However, the mechanism behind multilingual reasoning in LRMs is not yet fully explored. To investigate the issue, we fine-tune multilingual LRMs with Group-Relative Policy Optimization (GRPO) on translated versions of the GSM$8$K and SimpleRL-Zoo datasets in three different languages: Chinese, Korean, and Ukrainian. During training, we monitor both task accuracy and language consistency of the reasoning chains. Our experiments reveal three key findings: (i) GRPO rapidly amplifies pre-training language imbalances, leading to the erosion of low-resource languages within just a few hundred updates; (ii) language consistency reward mitigates this drift but does so at the expense of an almost 5 - 10 pp drop in accuracy. and (iii) the resulting language collapse is severely damaging and largely irreversible, as subsequent fine-tuning struggles to steer the model back toward its original target-language reasoning capabilities. Together, these findings point to a remarkable conclusion: \textit{not all languages are trained equally for reasoning}. Furthermore, our paper sheds light on the roles of reward shaping, data difficulty, and pre-training priors in eliciting multilingual reasoning.

Related papers

• Learn Globally, Speak Locally: Bridging the Gaps in Multilingual Reasoning [38.52080213211765]
  We introduce GeoFact-X, a geography-based multilingual factual reasoning benchmark with annotated reasoning traces in five languages.<n>We propose BRIDGE, a novel training method that guides supervised fine-tuning and test-time reinforcement learning.<n>Our results show that BRIDGE significantly enhances multilingual reasoning fidelity.
  arXiv  Detail & Related papers  (2025-07-07T19:04:36Z)
• When Models Reason in Your Language: Controlling Thinking Trace Language Comes at the Cost of Accuracy [9.021965237274244]
  Large Reasoning Models (LRMs) with thinking traces have shown strong performance on English reasoning tasks.<n>This capability is as important as answer accuracy for real world applications because users may find the reasoning trace useful for oversight only when it is expressed in their own language.<n>We evaluate two leading families of LRMs on our XReasoning benchmark and find that even the most advanced models often revert to English or produce fragmented reasoning in other languages.
  arXiv  Detail & Related papers  (2025-05-28T21:44:12Z)
• MMATH: A Multilingual Benchmark for Mathematical Reasoning [94.05289799605957]
  We introduce MMATH, a benchmark for multilingual complex reasoning spanning 374 high-quality math problems across 10 typologically diverse languages.<n>We observe that even advanced models like DeepSeek R1 exhibit substantial performance disparities across languages and suffer from a critical off-target issue-generating responses in unintended languages.<n>Our findings offer new insights and practical strategies for advancing the multilingual reasoning capabilities of large language models.
  arXiv  Detail & Related papers  (2025-05-25T12:47:39Z)
• Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models? [59.970391602080205]
  Despite multilingual training, LRMs tend to default to reasoning in high-resource languages at test time.<n>Cultural reasoning degrades performance on reasoning tasks but benefits cultural tasks, while safety evaluations exhibit language-specific behavior.
  arXiv  Detail & Related papers  (2025-05-23T02:46:18Z)
• When Less Language is More: Language-Reasoning Disentanglement Makes LLMs Better Multilingual Reasoners [111.50503126693444]
  We show that language-specific ablation consistently boosts multilingual reasoning performance.<n>Compared to post-training, our training-free ablation achieves comparable or superior results with minimal computational overhead.
  arXiv  Detail & Related papers  (2025-05-21T08:35:05Z)
• Crosslingual Reasoning through Test-Time Scaling [51.55526326294275]
  We find that scaling up inference compute for English-centric reasoning language models (RLMs) improves multilingual mathematical reasoning across many languages.<n>While English-centric RLM's CoTs are naturally predominantly English, they consistently follow a quote-and-think pattern to reason about quoted non-English inputs.<n>We observe poor out-of-domain reasoning generalization, in particular from STEM to cultural commonsense knowledge, even for English.
  arXiv  Detail & Related papers  (2025-05-08T16:50:06Z)
• Scaling Test-time Compute for Low-resource Languages: Multilingual Reasoning in LLMs [3.9530780161144667]
  We investigate the multilingual mechanism by which Large Language Models internally operate in a latent space biased toward their inherently dominant language.<n>We train models to generate the chain-of-thought (CoT) in English while outputting the final response in the target language, given input in the low-resource language.<n>Our experiments demonstrate that this approach, named English-Pivoted CoT Training, outperforms other baselines, with up to 28.33% improvement.
  arXiv  Detail & Related papers  (2025-04-02T16:58:36Z)
• Demystifying Multilingual Chain-of-Thought in Process Reward Modeling [71.12193680015622]
  We tackle the challenge of extending process reward models (PRMs) to multilingual settings.<n>We train multilingual PRMs on a dataset spanning seven languages, which is translated from English.<n>Our results highlight the sensitivity of multilingual PRMs to both the number of training languages and the volume of English data.
  arXiv  Detail & Related papers  (2025-02-18T09:11:44Z)
• AdaMCoT: Rethinking Cross-Lingual Factual Reasoning through Adaptive Multilingual Chain-of-Thought [40.16140566668239]
  We introduce AdaMCOT, a framework that enhances multilingual factual reasoning.<n>AdaMCOT dynamically routing thought processes in intermediary "thinking languages" before generating target-language responses.<n>Our evaluation demonstrates substantial improvements in both factual reasoning quality and cross-lingual consistency.
  arXiv  Detail & Related papers  (2025-01-27T15:48:57Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.