What we research.

Six research lines, one underlying question across all of them: how do we know this system is trustworthy? Below you'll find each front — questions, methods, sub-topics and recent work.

Trustworthy AI

"How do we know this model is trustworthy — before someone has to rely on it?"

The lab's chassis-line: we study the conditions under which a learned model can be deployed in real decisions. That involves quantifying uncertainty, auditing behavior against specification, calibrating probabilities, and exposing the failure modes a "100% accurate" model hides.

We work alongside all five other areas — robustness, fairness, language models, recommendation, intelligent systems — because trustworthiness is not a layer you bolt on later, it has to enter the design.

Methods Conformal prediction Calibration metrics (ECE, MCE) Audit harnesses Bayesian deep learning Selective prediction

01.a Calibration Probabilities that match actual frequencies — temperature, isotonic, conformal.

01.b Auditing Protocols to check production-model behavior against requirements.

01.c Uncertainty Epistemic and aleatoric quantification — abstention, deferral, decision under risk.

Recent work

2026 Lightweight post-training calibration for Portuguese LLMs NeurIPS
2025 Auditing credit classifiers under demographic shift FAccT
2025 Conformal prediction for top-k recommendation RecSys

Model robustness

Robustness

"Training is over — now the world shifts. Does the model hold?"

Models leave the lab with great accuracy and break in production for boring reasons: shifting distributions, deliberately manipulated inputs, fragile decision boundaries. We study how to measure, guarantee and recover robustness under those real pressures.

Scope spans from distribution shift in tabular and time-series settings to adversarial perturbations in language and vision models, with attention to when robustness is worth the trade-off against clean accuracy.

Methods Adversarial training Distributionally robust optimization Domain adaptation Test-time augmentation Certified defenses

02.a OOD & distribution shift Detection and adaptation when production data leaves training distribution.

02.b Adversarial robustness Defenses and auditing against malicious perturbations in vision, NLP and recsys.

Recent work

2026 Certified robustness for recommenders under injection attacks KDD
2025 OOD detection for clinical classification on Brazilian data BRACIS

Recommender systems

Recommender Systems

"How do we recommend when nobody chose anything yet — and how do we measure if the recommendation helped?"

The lab's historical line, with a strong methodological tradition. We work both ends of the problem: cold start (new user, new item, new domain) and diversity (when a giant catalog collapses to the same five items).

Crosses with fairness — recommending fairly for producers and consumers — and with robustness — recommending under injection attacks and feedback loops. Applications in personalized education, retail, and media.

Methods Matrix & tensor factorization Sequential / session-based Graph neural networks Contextual bandits Counterfactual evaluation

03.a Diversity Promoting diversity without destroying accuracy: measuring, optimizing, auditing the trade-off.

03.b Cold start Recommendation for users, items and domains with no history — meta-learning, side info, hybrid models.

Recent work

2026 Large-scale educational recommendation from weak signals WSDM
2025 Calibrated diversity for long-tail catalogs RecSys
2025 Counterfactual evaluation of production recommenders KDD

Algorithmic fairness

Fairness

"What is a fair algorithmic decision — and how do we audit it before it becomes policy?"

We treat algorithmic fairness as a property that has to be defined, measurable and auditable, not as an adjective. We operate on two complementary fronts: group parity (the aggregate snapshot of the system) and counterfactuals (what would change for this individual).

We work in sensitive domains — credit, health, education, platforms — always with the question: does this fairness metric reflect the real harm we're trying to prevent?

Methods Group fairness metrics Causal & counterfactual inference In-processing constraints Post-processing reweighing Disaggregated evaluation

04.a Group parity Demographic parity, equalized odds, equal opportunity — when each applies.

04.b Counterfactual fairness Individual-level fairness via causal models — what would change, and for whom.

Recent work

2026 Fairness in recommender systems: PhD dissertation PESC/COPPE
2025 Counterfactual reweighing for credit classifiers FAccT

Intelligent systems

Intelligent Systems

"When pure learning isn't enough — and when classical optimization needs learning to scale."

The lab's bridge line: here machine learning meets combinatorial optimization, graph search, simulation and classical AI. We build hybrid systems where each component sits in the right place — neural when it solves, symbolic when it must.

Applications in planning, logistics, resource allocation and simulation of complex domains, in collaboration with public and private partners.

Methods Combinatorial optimization Metaheuristics Reinforcement learning Neuro-symbolic Simulation-based learning

05.a Optimization Exact, heuristic and learned methods for combinatorial and continuous problems.

05.b Hybrid models Neural + symbolic, neural + simulator, neural + solver — where they compose, where they break.

Recent work

2025 Learning heuristics for routing under real-world constraints ENIAC
2025 Neuro-symbolic models for long-series diagnosis BRACIS

Language models

Language Models

"Language models that understand Brazil — and that know when they don't know."

We study LLMs with three commitments: real quality in Brazilian Portuguese, retrieval augmentation (don't invent when you can look it up) and alignment to what makes sense in the Brazilian sociotechnical context.

We build benchmarks, run human and automated evaluation, and study where calibration (area 01) and robustness (area 02) enter generative models — because a language model that decides is a language model that must be audited.

Methods Instruction tuning RAG (retrieval-augmented gen.) Preference optimization (DPO) LLM-as-judge eval PT-BR benchmark design

06.a PT-BR evaluation Benchmarks, metrics and protocols built for Brazilian Portuguese LLMs.

06.b Retrieval-augmented gen. RAG querying Brazilian official sources — legislation, public data, domain knowledge.

06.c Alignment Alignment to local context — not imported from English-language pipelines.

Recent work

2026 Four Brazilian benchmarks for LLM evaluation ACL
2026 RAG over Brazilian federal legislation NeurIPS
2025 Human evaluation of PT-BR assistants — open protocol arXiv

How we research

Principles that cut across all six areas.

01 · Open

Open science by default

Preprints on arXiv, code in public repositories, reproducible benchmarks. A result no one can audit doesn't count.

02 · Anchored

Brazilian problem, frontier method

We don't import English-language pipelines and rename them. We build benchmarks, datasets and protocols for the local context.

03 · Cross-cut

The six lines talk to each other

Calibration shows up in LLMs. Fairness shows up in recommenders. Robustness shows up in intelligent systems. These are not silos.

04 · Trained

Research that trains researchers

Each line has PESC/COPPE master's and PhD students leading the work — not as assistants, but as authors.