MILE-RefHumEval: A Reference-Free, Multi-Independent LLM Framework for Human-Aligned Evaluation

ACL ARR 2025 May Submission2973 Authors

Abstract: We present MILE-RefHumEval, a novel reference-free framework for evaluating Large Language Models (LLMs) without the need for ground-truth annotations or coordination among evaluators. It leverages multiple independently prompted LLMs and a 12-point human-aligned schema to generate nuanced, high-quality assessments. The framework demonstrates strong alignment with human judgment and consistently outperforms prior approaches. Importantly, it delivers these gains with substantially reduced computational overhead, making it a scalable, efficient, and human-aligned solution for evaluating LLMs in open-ended, real-world tasks.

Paper Type: Short

Research Area: Generation

Research Area Keywords: Automatic Evaluation, document-level extraction, zero/few-shot extraction, LLM/AI agents

Contribution Types: Position papers

Languages Studied: English

Submission Number: 2973