LH-Bench: Evaluating AI on Subjective Enterprise Tasks

Researchers have introduced LH-Bench, a new evaluation framework for assessing large language models on subjective, long-horizon enterprise tasks. Unlike traditional benchmarks that focus on objectively verifiable problems like math and programming, LH-Bench addresses the complexity of real-world work where success depends on organizational goals, user intent, and intermediate artifacts. The framework employs three pillars: expert-grounded rubrics providing domain context for LLM judges, curated ground-truth artifacts enabling stepwise reward signals, and pairwise human preference evaluation for convergent validation. The study demonstrates that domain-authored rubrics yield substantially more reliable evaluation. The work is detailed in arXiv:2603.22744.