Senior Research Scientist · Sword Health
José Pombal
- Google Scholar
- GitHub
- j.pombal (at) swordhealth.com
I am a Senior Research Scientist at Sword Health, and a PhD student at Instituto Superior Técnico and Instituto de Telecomunicações in Lisbon, advised by Prof. André Martins and Ricardo Rei. I’m currently working on building frontier LLMs for healthcare with a focus on applications for mental health, and I’m particularly interested in the automatic evaluation of such models. Previously, I worked on the development and evaluation of multilingual LLMs at Unbabel, and fairness-aware ML for fraud detection at Feedzai.
News
- Oct 2026 Released MindEval2, which pairs interpretable rubrics developed with clinical psychologists with a more realistic patient simulator that keeps memory across sessions. Code and data.
- Apr 2026 Self-preference bias in rubric-based evaluation, showing that LLM judges favor their own outputs even when rubrics are objectively verifiable, accepted at COLM 2026. Code.
- Nov 2025 Released MindEval, a framework built with licensed clinical psychologists for evaluating LLMs in realistic, multi-turn mental health conversations, accepted at the DAIH workshop @ COLM 2026. Code and data are on GitHub, and there’s a blog post with a video.
- Jul 2025 Joined Sword Health as a Senior Research Scientist.
- 2023–2025 Joined Unbabel as a Research Scientist in March 2023, working on post-training and automatic evaluation of multilingual LLMs, with a focus on machine translation. Papers include Tower (COLM 2024 spotlight), Tower v2 and TowerChat (winners of three WMT24 shared tasks), Tower+ (ACL 2026), M-Prometheus and Zero-shot Benchmarking (COLM 2025), and TACL papers on metric interference and context-aware translation. Open models: Tower, Tower+, M-Prometheus.
Selected publications
* Equal contribution. Full list on Google Scholar.