Translation cascades for reasoning translate queries to English, reason, and translate back, but this process is structurally lossy due to information discard at each stage. The authors propose a context-aware translation cascade that preserves the original question, translated query, and reasoning trace to mitigate these losses.
- The intervention provides the original question, English translated question, and reasoning trace to the final translation module without requiring training.
- Evaluation covers nine multilingual benchmarks, three backbone models, and 285 high-, mid-, and low-resource languages.
- Strong gains in open-ended generation were demonstrated across all models and resource regimes.
- The original language question was found to carry most of the beneficial context for the final translation.
The study emphasizes designing better information flow in machine translation cascades to mitigate error propagation, offering a simple strategy to preserve the original user question until the end of the pipeline.