E01¶


Research	R0027 — Multilingual prompt engineering challenges
Run	2026-03-26
Query	Q001
Source	SRC02
Evidence	SRC02-E01
Type	Statistical

Selective pre-translation outperforms both full translation and native inference

URL: https://arxiv.org/html/2502.09331v1

Extract¶

"Selective pre-translation consistently outperforms both pre-translation and direct inference in the source language" across 35 languages, 4 tasks, and multiple models. The study tested 24 prompt configurations per language/task combination, varying which components (instruction, context, examples, output) were in English vs native language. Task-specific patterns emerged: extractive tasks perform better with source language context, while generative tasks benefit from English output.

Relevance to Hypotheses¶

Hypothesis	Relationship	Strength
H1	Supports	Demonstrates measurable performance differences between prompt language strategies
H2	Contradicts	Clear, consistent performance gaps found across all tested configurations
H3	Supports	The optimal strategy varies by task type and language — strongly conditional

Context¶

This is one of the most methodologically rigorous studies found. The 24 configurations per language/task allow fine-grained analysis of which prompt components benefit from translation.