What’s Next for Brazil’s AI Community: Key Insights from BRACIS 2025
What’s Next for Brazil’s AI Community: Key Insights from BRACIS 2025
Fortaleza brought the heat in more ways than one. At the 35th Brazilian Conference on Intelligent Systems (BRACIS 2025), co-located with the Brazilian Symposium on Databases, roughly a thousand researchers, students, and industry builders gathered to discuss the future of AI, particularly in Portuguese-speaking markets. This event focused on emerging trends, practical applications, and the challenges that lie ahead for the Brazilian AI community.
The Shift to Excellence in Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation (RAG) combines information retrieval with language generation, helping AI systems use existing data efficiently. At BRACIS 2025, discussions transitioned from "does it work?" to "how can we make it exceptional?" Presenters highlighted innovative techniques like smarter example selection for few-shot learning and multi-objective prompt optimization.
For instance, improvements in self-supervised fine-tuning tailored for low-resource Portuguese environments received significant attention. These advancements allow for high-quality outputs without extensive labeling costs. Ultimately, this evaluation discipline enhances the capability of smaller models, benefiting businesses through lower latency and operational costs.
Maturation of Legal AI in Brazilian Portuguese
Legal AI is gaining momentum, and BRACIS showcased this evolution. This year marked an inflection point, with a legal-search hackathon and new datasets specifically for Brazilian personal income tax queries. Projects that involved classifying Brazilian Supreme Court (STF) documents using pseudo-labeling exemplified community efforts to bridge gaps in legal frameworks.
For example, utilizing Layer-wise Relevance Propagation (LoRA) for legal named entity recognition (NER) displayed potential. By training a domain-specific Brazilian legal model informed by reputable sources, the community is laying the groundwork for dependable systems that meet compliance requirements.
The Resurgence of Meta-Learning and Bio-Inspired Algorithms
Meta-learning focuses on developing models that adapt quickly with minimal data as tasks or domains shift. At BRACIS, meta-learning sessions emphasized practical applications, while bio-inspired algorithms provided alternative solutions for optimizing processes where traditional transformers may falter.
For instance, hybrid approaches that combine these methodologies could efficiently tackle optimization challenges in various sectors. Knowing that not all problems yield to singular approaches, attendees acknowledged that embracing adaptability could yield higher performance and innovation in AI applications.
Memorable Sessions That Stood Out
Low-resource and small language models received significant attention. Four papers analyzed techniques from few-shot named entity recognition to text-to-SQL conversion using small language models (SLMs). These findings present valuable insights applicable in various settings, emphasizing the importance of tailored example selection in workflows.
Another notable area of discussion was the practical applications of NLP in legal settings. Presentations showcased pseudo-labeling techniques for classifying STF documents, LoRA implementations targeting legal NER, and datasets dedicated to Brazilian tax law Q&A. The key takeaway was the need for rigor in the design process, ensuring outputs are both precise and sourced from trusted materials.
Engaging Conversations at the Booth
The Thomson Reuters Labs booth became a hub for networking, attracting researchers, technologists, and students from various regions of Brazil. Participants expressed a keen interest in applying AI developments to real-world challenges, particularly in legal and tax-related contexts. This engagement demonstrates a growing synergy between academia and industry, paving the way for innovative collaborations.
Brazil’s AI landscape is evolving dynamically. The focus at BRACIS was not merely on high-profile demonstrations but on building reliable systems capable of meeting rigorous standards set by courts, regulators, and clients. These foundational elements will undoubtedly contribute to the sustainability of AI initiatives across the country.
BRACIS 2025 exemplified a forward-thinking environment where the future of AI is defined by multilingual capabilities, domain-specific applications, and a commitment to rigorous evaluation methods. As the community grows, the commitment to integrating practical solutions and real-world constraints into AI design will continue to be crucial.
Obrigado, Fortaleza. We will be back.