BLUEX v2: Benchmarking LLMs on Open-Ended Questions from Brazilian University Entrance Exams

BLUEX v2 introduces a new benchmark for evaluating Large Language Models (LLMs) on open-ended, discursive tasks derived from the second-phase entrance exams of Brazil's UNICAMP and USP, covering exam years 2022-2025. The benchmark includes 395 questions and 919 graded subquestions, with 55.7% containing images, and evaluates 21 state-of-the-art LLMs using an LLM-as-a-judge protocol, revealing a performance spread from 4.18 to 9.10 on a 0-10 scale. This resource is significant for practitioners as it addresses the gap in Portuguese-language LLM assessment, particularly for complex reasoning tasks, and provides a publicly available dataset and evaluation framework.