Multilingual Reasoning Cascades Need More Context

2026-06-25

AI Summary

This work addresses the degradation of answer quality in multilingual reasoning caused by traditional translation cascades, which suffer from progressive information loss and fail to preserve cultural context, stylistic nuances, and ambiguity cues. The authors propose a training-free, context-aware translation cascade that integrates the original query, its English translation, and intermediate reasoning traces directly into the final reasoning stage to retain critical contextual information. This approach provides the first systematic validation of the pivotal role played by the source-language question within cascade pipelines and introduces a streamlined information-flow mechanism that routes the original query straight to the pipeline’s end. Built upon existing large language models and machine translation systems, the resulting context-enhanced architecture significantly improves open-ended generation performance across nine benchmarks, three backbone models, and 285 high-, medium-, and low-resource languages, with the most pronounced gains observed in low-resource settings.

Abstract

Translation cascades for reasoning translate the query from another language to English, reason in English, and translate the answer back to the original language. This is a competitive approach to multilingual reasoning, but structurally lossy, since each stage discards information later stages may need, including cues for cultural grounding, register, and disambiguation. We examine the benefits of a simple and training-free intervention: a context-aware translation cascade, which additionally provides the original question, the English translated question, and the reasoning trace to the context of the final translation module. We evaluate gains across nine multilingual benchmarks including various task types, three backbone models, and 285 high-, mid-, and low-resource languages, and demonstrate strong gains for open-ended generation across models and resource regimes. We show that the original language question carries most of the beneficial context. Our study emphasizes the need to better design information flow in machine translation cascades for mitigating error propagation, and provides a simple and actionable default strategy: preserve the original user question until the end of the pipeline.

Problem

• multilingual reasoning
• translation cascades
• context loss
• error propagation
• machine translation

Innovation

• multilingual reasoning
• translation cascades
• context-aware translation
• error propagation mitigation
• original question preservation