João Miguel Coelho
My research interests lie in the areas of Machine Learning, Natural Language Processing, and Information Retrieval.
Currently a PhD Candidate at IST (DEEC) and CMU (LTI), under the dual degree CMU-Portugal program, supervised by Bruno Martins, João Magalhães, and Chenyan Xiong.
My current research focuses on dense retrieval and deep search systems, including training LLM search agents with reinforcement learning, and building infrastructure for deep research. I spent summer 2025 as an Applied Scientist Intern at Amazon. I worked with the AWS Connect team on agentic search for contact centers.
Recent Work
| Jun 2026 |
Effective Reinforcement Learning for Agentic Search by Recycling Zero-Variance Queries During Training Returning zero-variance queries to the training pool for later resampling makes GRPO training of search agents more effective, with a 1.7B agent matching prior systems of up to 7B parameters. |
|---|---|
| Jun 2026 |
Beyond Parallel Sampling: Diverse Query Initialization for Agentic Search DivInit is a training-free method that picks diverse first-turn queries for parallel search trajectories, reducing the redundancy that limits standard parallel sampling. |
Selected Publications
-
Dwell in the Beginning: How Language Models Embed Long Documents for Dense RetrievalProceedings of the Annual Meeting of the Association for Computational Linguistics 2024
-
DeepResearchGym: A Free, Transparent, and Reproducible Sandbox for Deep ResearchProceedings of the ACM SIGIR Conference on Innovative Concepts and Theories in Information Retrieval 2026