---
title: "An Actionable Diagnosis of Multilingual, Multi-Agent Planning Failures: TART"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.03735v1"
permalink: "https://feed.astralor.com/items/itm_30c565194a8d490d5dd4fead"
published: "2026-08-04T14:29:14.000Z"
score: 80
channel: "ai"
tags: ["GAIA","TART","Multi-agent Systems","Multilingual"]
---

# An Actionable Diagnosis of Multilingual, Multi-Agent Planning Failures: TART

## AI commentary

> This paper is valuable because it moves multilingual agent failures from vague language-resource claims to concrete planning-grounding issues, and offers a directly usable mitigation. Teams building agent infrastructure or low-resource language products should try the TART approach immediately.

## At a glance

Multilingual multi-agent systems degrade substantially beyond English, yet prior work rarely identifies how task-critical information is lost when user requests are converted into executable plans. This paper treats the planner as the request-to-action interface, derives an actionable taxonomy of planning-grounding failures from failed executions, and introduces TART, which makes key taxonomy aspects explicit to the planner and downstream sub-agents. Across multiple languages, three LLM backbones, two datasets, and two agentic configurations, TART consistently improves performance, raising state-of-the-art accuracy by 5.6 percentage points on multilingual GAIA averaged over eleven languages.

## Key points

- Derives an actionable taxonomy of planning-grounding failures from real-world failed executions, showing these failures increase as language resources decline.
- Introduces TART (Taxonomy-Guided Actionable Representation) to make key taxonomy aspects explicit to planners and downstream sub-agents.
- Consistently improves performance across multiple languages, three LLM backbones, two datasets, and two agentic configurations.
- Raises state-of-the-art accuracy by 5.6 percentage points on multilingual GAIA across eleven languages.

## Full text

Multilingual multi-agent systems exhibit substantial degradation beyond English, yet prior work rarely identifies how task-critical information is lost when user requests are converted into executable plans. We study the planner in a multi-agent system as the request-to-action interface and derive an actionable taxonomy of planning-grounding failures from failed real-world task executions. LLM-based analysis shows that these failures constitute an increasing share of unsuccessful executions as language-resource availability declines, with the strongest effects in low-resource languages. To test whether the taxonomy supports mitigation, we introduce TART, Taxonomy-Guided Actionable Representation, that makes the taxonomy's key aspects explicit to the planner and downstream sub-agents. Across multiple languages, three LLM backbones, two datasets, and two agentic configurations, TART consistently improves performance. On multilingual GAIA, it raises a state-of-the-art system's accuracy by 5.6 percentage points averaged across eleven languages spanning low- to high-resource settings.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_30c565194a8d490d5dd4fead