Microsoft Maps Multilingual Agent Planning Failures, Lifts Accuracy 5.6%

A nine-author team from Microsoft Research, University of Edinburgh, and industry partners published the first structured failure taxonomy for multilingual agent pipelines. The failure mode doesn't live at the LLM reasoning layer but at the planner — the component that converts user intent into executable sub-task instructions. The paper, "An Actionable Diagnosis of Multilingual, Multi-Agent Planning Failures," appeared on arXiv on August 4, 2026, with a mitigation technique called TART. TART adds 5.6 percentage points of average accuracy on multilingual GAIA across eleven languages.

When a non-English request hits a multi-agent pipeline, the planner parses it and emits a structured plan: tool calls, sub-agent instructions, API parameters. If the planner mis-grounds intent at this step, downstream agents execute faithfully on the wrong plan. No exception fires. The agent completes. The output is wrong. The authors term this the request-to-action interface. Planning-grounding failures grow as a share of total failures the further you move from high-resource languages like English or German.

The taxonomy stems from failed real-world executions, not synthetic adversarial inputs. Researchers classified failures across execution traces, bucketing them into categories that planners can recognize and act on. Each failure type carries a diagnostic signal, making the taxonomy "actionable" rather than post-hoc. In low-resource settings, planning-grounding failures account for the largest and fastest-growing share of unsuccessful runs.

TART makes the taxonomy's failure categories explicit in the planner's prompt and instructions to downstream agents. Rather than retraining, it scaffolds both layers to watch for specific failure patterns. The evaluation spans three LLM backbones, two datasets (including multilingual GAIA), and two agentic configurations. TART lifts the state-of-the-art GAIA system by 5.6 percentage points across eleven languages, low- to high-resource. The gain holds across all three models — the improvement doesn't depend on any single provider's multilingual training.

Separate LILT research using MultiChallenge found that Instruction Retention and Inference Memory drop 3–7% across non-English languages even for frontier models. Reliable Version Editing collapses to 32–37% for Arabic and Korean versus much higher English baselines. The analysis attributes 70–80% of the non-English gap to fundamental model limitations — tokenizer inefficiency and English-centric reasoning — not data artifacts. The MAPS benchmark shows agentic systems directly inherit language-dependent failure modes, adding incorrect tool invocations and degraded confidence calibration on top of accuracy drops.

Prior work identified language-dependent degradation. This paper adds a planner-specific failure map. Knowing your agent degrades on Japanese isn't actionable. Knowing the planner mis-grounds object references from inflected noun phrases into tool parameters is. TART's prompt approach won't close the full gap — low-resource degradation persists even with the intervention — but it provides a diagnostic skeleton teams can extend, log, and triage. Taxonomy categories become first-class fields in execution traces, not post-incident inferences.

If you route non-English traffic through a multi-agent pipeline, instrument planner output as a first-class signal. Plan correctness — whether the emitted plan matches grounded user intent in the source language — is a separate failure mode from agent execution failure. Conflating them masks the dominant error source in non-English deployments.