Researchers reveal why multilingual AI agents keep failing

A team of 9 researchers from institutions including Google, NVIDIA, and Hebrew University have published a groundbreaking paper identifying a critical weakness in multilingual AI systems: planning failures that disproportionately affect non-English languages. The 22-page study titled 'An Actionable Diagnosis of Multilingual, Multi-Agent Planning Failures' reveals that task-critical information is systematically lost when user requests are translated into executable plans, with failure rates increasing as language resources decline.

The researchers developed TART (Taxonomy-Guided Actionable Representation), a framework that makes planning-grounding failures explicit to both planners and sub-agents. Tested across multiple languages, three LLM backbones, two datasets, and two agentic configurations, TART consistently improved performance. On the multilingual GAIA benchmark, it boosted a state-of-the-art system's accuracy by 5.6 percentage points averaged across 11 languages spanning from low- to high-resource settings, with the largest gains in languages with limited training data.