Seeing the Whole Elephant: A Benchmark for Failure Attribution in LLM-based Multi-Agent Systems

Technical Summary: Seeing the Whole Elephant

Title: Seeing the Whole Elephant: A Benchmark for Failure Attribution in LLM-based Multi-Agent Systems
Authors: Mengzhuo Chen, Junjie Wang, Fangwen Mu, et al.

1. Problem Statement

As Large Language Model (LLM)-based Multi-Agent Systems (MASs) become more complex, identifying the root cause of a failure—failure attribution—becomes increasingly difficult. Attribution requires identifying both the responsible agent and the decisive failure step (the earliest point where failure becomes inevitable).

The authors identify a critical gap in existing research: current benchmarks (such as Who&When) rely on partial observability, providing only the agents' outputs. In real-world software debugging, developers have access to full observability, including task instructions, prompts, intermediate messages, tool logs, and environment states. Relying only on outputs obscures the "why" behind a failure, as the input context is often the deciding factor in an agent's reasoning error.

2. Methodology

The paper introduces TraceElephant, a benchmark designed for developer-facing failure attribution under full execution observability.

• Benchmark Construction:

  • Data Sources: Traces are collected from three diverse agentic architectures: Captain-Agent (dynamic team assembly), Magentic-One (fixed-role orchestration), and SWE-Agent (single-agent tool-centric scaffold).
  • Trace Schema: Unlike previous works, TraceElephant records a complete "execution narrative," including trace metadata (system architecture, configurations), step-level inputs (task instructions, context, prompts), and step-level outputs (LLM responses, raw tool/environment logs).
  • Reproducibility: Each trace is accompanied by a reproducible execution environment, allowing for "dynamic" debugging.
  • Annotation: Labels for the responsible agent and decisive step were generated through a multi-round expert consensus process to ensure high reliability.

• Evaluation Framework:
  The authors evaluate attribution techniques under two configurations:

  • Static: Using the complete recorded trace (metadata + inputs + outputs).
  • Dynamic: Using the static trace plus a replayable environment to perform counterfactual probing (e.g., "If I change this input, does the failure still occur?").

3. Key Contributions

1. TraceElephant Benchmark: The first benchmark for MAS failure attribution that provides full execution traces and reproducible environments.
2. Paradigm Shift: Moves the field from "black-box" (output-only) evaluation to "developer-centric" (full-trace) evaluation.
3. Systematic Analysis: Provides a comprehensive study of how observability, backbone LLM capability, and system architecture affect attribution accuracy.

4. Key Results

• The Importance of Observability: Full traces improve step-level attribution accuracy by up to 76% compared to output-only traces. This confirms that missing inputs are the primary cause of ambiguity in failure localization.
• Dynamic vs. Static: The Dynamic Agentic approach (using re-execution and counterfactual checks) achieves the highest performance, improving step-level accuracy by 10% over static methods.
• Performance Trends:
  • Agent-level accuracy averaged 65.9%, while step-level accuracy averaged 30.3%.
  • Backbone Models: Advanced models like Claude-3.5-Sonnet and DeepSeek-R1 significantly outperform smaller models, highlighting that attribution is a high-level reasoning task.
  • Failure Patterns: Agents interacting with external environments (web browsing, code editing) are the most frequent sources of failure (>50%).
• Complexity of Steps: Early-phase failures (e.g., incorrect initial planning) are significantly harder to attribute than late-phase failures, which often present clearer, more localized signals.

5. Significance

This work provides a foundational tool for the development of more transparent and debuggable MASs. By shifting the focus to full observability, the authors provide a realistic playground for researchers to develop advanced attribution techniques, such as:

• Architecture-aware attribution (tailoring methods to specific MAS designs).
• Automated debugging tools that integrate real-time trace capture and interactive hypothesis testing.
• Specialized attribution models fine-tuned on rich, structured execution narratives.