An Actionable Diagnosis of Multilingual, Multi-Agent Planning Failures
3.60T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2608.03735.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryThe paper studies how multilingual multi-agent systems degrade at the planning stage, derives a taxonomy of planning-grounding failures from real failed executions, and introduces TART, which improves accuracy by 5.6 percentage points on multilingual GAIA across eleven languages and multiple LLM backbones and configurations.
Why it mattersPinpoints a specific, under-studied failure mode in multilingual agent pipelines and offers a tested mitigation with measurable gains across languages and models.
Cited by
No citations on record.