LILT Unveils AURORA to Benchmark Multilingual AI Agent Performance

The shift toward agentic AI means that language errors no longer result in simple translation mishaps but in faulty real-world decision-making. Spence Green, CEO of LILT, warned that companies choosing models based solely on English-centric scoreboards risk operational failure, as these systems often falter when tasked with complex, localized workflows.

AURORA evaluates models through a suite of benchmarks curated by native-language experts, covering sectors including software development, retail, and banking. The platform tracks performance across categories such as multi-turn customer support and long-context instruction following. Early analysis from LILT’s research team reveals significant quality discrepancies: while GPT 5.5 leads in Spanish coding tasks, Claude Opus 5.5 excels in Japanese, and Muse Spark 1.3 outperforms peers in Serbian. This data suggests that model efficacy is highly variable depending on the target language and specific domain requirements.