Multilingual Reasoning Cascades Need More Context

A new study from arXivLabs finds that multilingual reasoning systems that translate queries into English, reason, and then translate back are structurally lossy, discarding information needed later in the pipeline [1][2]. The researchers propose a context-aware cascade that preserves the original question until the final step, demonstrating gains across 285 languages and nine benchmarks [2]. The standard translation cascade for reasoning translates a query from another language to English, performs the reasoning in English, and then translates the answer back to the original language [1][2]. While competitive, this approach is structurally lossy because each stage discards information that later stages may need, including cues for cultural grounding, register, and disambiguation [2]. The paper, posted on the open-access repository arXiv on June 25, 2026, examines a training-free intervention: a context-aware translation cascade [2][5]. This method provides the original question, the English translated question, and the reasoning trace as context for the final translation module [1][2]. The authors evaluated the approach across nine multilingual benchmarks covering various task types, three backbone models, and 285 high-, mid-, and low-resource languages [1][2]. The results showed strong gains for open-ended generation across models and resource regimes [2]. The original language question carried most of the beneficial context, leading the researchers to recommend preserving the original user question until the end of the pipeline as a default strategy [1][2]. The study emphasizes the need to better design information flow in machine translation cascades to mitigate error propagation [1][2]. arXiv, which began on August 14, 1991, now receives about 24,000 article submissions per month and hosts preprints that are moderated but not peer-reviewed [5]. The repository passed the two-million-article milestone by the end of 2021 [5]. The paper's findings contribute to ongoing work on how groups of artificial agents coordinate information and tasks, a domain related to broader research on collective intelligence [4]. Collective intelligence refers to the emergent ability of groups to solve problems more effectively than individuals through cooperation or aggregation of diverse information [4]. The arXivLabs framework, which supported the study, allows collaborators to develop and share new features directly on the arXiv website under values of openness, community, and user data privacy [1].