Retrieval-Augmented Generation (RAG) is proposed to reduce hallucination and improve grounding in clinical language models, yet its effectiveness across different levels of clinical reasoning remains unclear. We conducted a controlled evaluation of medication-related question answering in Portuguese using over 7,000 Brazilian regulatory drug leaflets and a complementary clinical benchmark derived from national medical licensing examinations (Revalida and Fuvest). Retrieval substantially improved factual recall and clinical coherence in medication-specific queries, increasing F1 from 0.276 to 0.412. However, naive retrieval did not consistently improve complex clinical reasoning and sometimes reduced accuracy compared to a parametric-only baseline. We identify retrieval-induced anchoring bias, where partially relevant evidence shifts model decisions toward clinically incorrect conclusions. Critique-based and adaptive retrieval mitigated this effect and achieved the highest clinical benchmark accuracy (54.25%). Clinically grounded evaluation dimensions revealed safety-relevant differences beyond traditional NLP metrics. These results show that retrieval augmentation is effective in regulatory settings but requires adaptive control for higher-level clinical reasoning.

Evaluating open-ended text generation in large language models remains challenging, particularly for non-English languages. We introduce EduBench, a comprehensive Portuguese-language benchmark comprising 3,149 discursive questions from Brazilian university entrance examinations spanning 2015–2025. Unlike multiple-choice or extractive QA benchmarks, EduBench requires extended, argumentative responses across diverse domains, including Humanities, Exact and Natural Sciences, and Languages. Each question includes expert-curated reference answers from official sources, rich metadata, and automated image descriptions to support text-only evaluation. We establish baseline results using nine contemporary models, ranging from 4B-parameter SLMs to state-of-the-art reasoning-capable LLMs, and evaluate them using complementary metrics (BLEU, BERTScore, G-Eval). Our results reveal substantial metric disagreement and highlight the complexity of assessing discursive generation, with models achieving 54–71% alignment with expert answers. We release EduBench publicly to support research on Portuguese NLP and open-ended generation evaluation.