LMUnit: Fine-grained Evaluation with Natural Language Unit Tests

TL;DR

This paper introduces natural language unit tests and LMUnit, a scoring model that decomposes language model response evaluation into explicit criteria, improving agreement and effectiveness in model assessment.

Contribution

The paper proposes a novel paradigm of natural language unit tests and a unified scoring model, enhancing evaluation accuracy and consistency for language models.

Findings

Improved inter-annotator agreement in evaluations.

State-of-the-art performance on FLASK and BigGenBench benchmarks.

Competitive results on RewardBench.

Abstract

As language models become integral to critical workflows, assessing their behavior remains a fundamental challenge -- human evaluation is costly and noisy, while automated metrics provide only coarse, difficult-to-interpret signals. We introduce natural language unit tests, a paradigm that decomposes response quality into explicit, testable criteria, along with a unified scoring model, LMUnit, which combines multi-objective training across preferences, direct ratings, and natural language rationales. Through controlled human studies, we show this paradigm significantly improves inter-annotator agreement and enables more effective LLM development workflows. LMUnit achieves state-of-the-art performance on evaluation benchmarks (FLASK, BigGenBench) and competitive results on RewardBench. These results validate both our proposed paradigm and scoring model, suggesting a promising path forward for language model evaluation and development.