Auto-Eval Judge: Towards a General Agentic Framework for Task Completion Evaluation

TL;DR

This paper introduces a modular, human-like evaluation framework for assessing agent task completion that considers step-by-step reasoning, outperforming existing methods in aligning with human judgments across multiple benchmarks.

Contribution

The paper presents a novel, domain-independent evaluation framework that decomposes tasks into sub-tasks and validates each step, improving alignment with human evaluations.

Findings

Achieves 4.76% higher alignment accuracy on GAIA

Achieves 10.52% higher alignment accuracy on BigCodeBench

Outperforms GPT-4-based LLM-as-a-Judge baseline

Abstract

The increasing adoption of foundation models as agents across diverse domains necessitates a robust evaluation framework. Current methods, such as LLM-as-a-Judge, focus only on final outputs, overlooking the step-by-step reasoning that drives agentic decision-making. Meanwhile, existing Agent-as-a-Judge systems, where one agent evaluates another's task completion, are typically designed for narrow, domain-specific settings. To address this gap, we propose a generalizable, modular framework for evaluating agent task completion independent of the task domain. The framework emulates human-like evaluation by decomposing tasks into sub-tasks and validating each step using available information, such as the agent's output and reasoning. Each module contributes to a specific aspect of the evaluation process, and their outputs are aggregated to produce a final verdict on task completion. We validate our framework by evaluating the Magentic-One Actor Agent on two benchmarks, GAIA and BigCodeBench. Our Judge Agent predicts task success with closer agreement to human evaluations, achieving 4.76% and 10.52% higher alignment accuracy, respectively, compared to the GPT-4o based LLM-as-a-Judge baseline. This demonstrates the potential of our proposed general-purpose evaluation framework.