OneMillion-Bench: Bilingual Expert-Level Language Agent Benchmark

Onemillion Bench is a bilingual (English/Chinese) expert-level benchmark containing 400 entries across five professional domains, released by humanlaya-data-lab in March 2026. It utilizes weighted rubric-based grading criteria to evaluate language agents on analytical reasoning and instruction following within specialized fields.

Use Cases

• Evaluating language agent performance in legal reasoning using domain-specific rubrics
• Benchmarking bilingual text generation accuracy in the healthcare and medicine domain
• Assessing instruction following for complex financial and economic analysis tasks

Strengths

• 400 expert-level entries
• Bilingual support for English and Chinese
• Weighted rubric-based grading for fine-grained evaluation
• Covers 5 distinct professional domains including Law and Healthcare

Limitations

• Small sample size of 400 total records
• Domain-specific focus may not reflect performance on general-purpose tasks
• Lack of column-level metadata in the initial documentation

Provenance

Source

humanlaya-data-lab (arXiv:2603.07980)

Collection Method

Expert-annotated

Freshness

Last updated March 2026.

Geography

Global (Bilingual English/Chinese)