VLegal-Bench Introduces First Comprehensive Benchmark for Vietnamese Legal AI

A new benchmark called VLegal-Bench has been created to evaluate how effectively large language models handle Vietnamese legal reasoning. This tool addresses the unique challenges posed by Vietnam's complex and frequently updated legal system. The benchmark contains 10,450 carefully annotated samples, each validated by legal experts against authoritative documents. Its design is informed by Bloom's cognitive taxonomy, structuring tasks to reflect real-world legal application scenarios. The development responds to the growing use of artificial intelligence in legal domains, where assessing model performance on specific national legislation has been difficult. This work was documented in the arXiv preprint 2512.14554v5.