Catarina G. Belem

Catarina G. Belem

PhD Student @ UCI

CV

A PDF version of my CV is available here.

Education

  • Ph.D., Computer Science, University of California, Irvine, 2021–Present
    • Advisors: Sameer Singh, Padhraic Smyth
    • Expected Graduation: November 2026
  • M.Sc., Computer Science and Engineering, Instituto Superior Técnico, University of Lisbon, 2016–2019
    • Thesis: Optimization of Time-Consuming Objective Functions: Derivative-Free Approaches and their Application in Architecture
  • B.Sc., Computer Science and Engineering, Instituto Superior Técnico, University of Lisbon, 2013–2016

Research Experience

  • Applied Research Intern — Apple (Summer 2026)
    • Implemented and experimentally evaluated approaches to improve reasoning efficiency in world-knowledge QA, exploring chain-of-thought reasoning, SFT, GRPO, and RLVR within a JAX-based research codebase.
  • Applied Research Intern — Capital One, New York (Jun–Sep 2025)
    • Investigated controllability and style transfer of LLMs for readable and accurate outputs.
    • Proposed and implemented a reinforcement learning customization approach (GRPO via TRL).
    • Authored 2 papers: workshop paper at EMNLP 2025 TSAR; working towards a conference paper on RL-based readability/accuracy alignment.
  • Research Intern — Megagon Labs, Mountain View (Jun–Sep 2024)
    • Analyzed hallucinations in multi-document summarization across 5 popular LLMs.
    • Proposed a taxonomy of error types through large-scale human annotation; evaluated adversarial robustness and mitigation strategies.
    • Resulted in a publication at NAACL 2025 Findings.
  • Graduate Student Researcher — University of California, Irvine (2021–Present)
    • Uncertainty in LLMs: Studying LLM calibration and linguistic uncertainty and its effects on human-AI decision-making and alignment with human perceptions.
    • Constrained Decoding: Proposed a method leveraging an attribute verifier’s gradient to efficiently steer LM generations by reweighting the next-token distribution.
    • Fair NLP: Developed an evaluation benchmark to uncover gender bias in non-stereotypical contexts (ICLR 2024).
    • NLG Evaluation: Applied PEFT and ICL for automatic evaluation of generative LLMs with few labeled examples.
  • Research Data Scientist — Feedzai, Lisbon, Portugal (2019–2021)
    • Algorithmic Fairness: Developed bias mitigation methods for model selection and training via constrained optimization and hyperparameter selection. Authored 2 patents and 2 top-tier papers (ICDM 2021, ICLR 2023).
    • Explainable AI: Evaluated the impact of explanations on human decision-making. Implemented concept-based explanations for fraud detection. Authored 3 patents and 3 papers (NeurIPS’20 WS, FAccT’21, ICLR’21 WS).
  • Graduate Student Researcher — INESC-ID, Lisbon, Portugal (2017–2019)
    • Investigated gradient-free methods (genetic algorithms, ML) for multi-objective optimization of time-consuming objective functions in architectural design.

Publications

Talks

Teaching

Skills

Awards & Fellowships

Service

Community Involvement