Information retrieval · RAG evaluation · reproducible ML
Gioia Zheng
Retrieval Diagnostics · RAG Evaluation · Failure Analysis · Reproducible ML
I study where retrieval-augmented generation fails, how retrieval and generation errors interact, and how evaluation can make those failures measurable. I build reproducible systems to turn these questions into testable evidence.
- LOC Rome, IT
- EDU B.Sc. ACSAI · Sapienza
- DOMAIN IR · RAG evaluation · reproducible ML
- CODE open source
Featured research projectAll projects
rag-observatory
activeResearch prototype for trace-based RAG observability and failure analysis.
Real traces40 SciFact
Controlled queries20 fixed
Automated tests107
Recent writingAll notes
When Self-Play Q-Learning Looks Robust but Remains Exploitable
2026-09-05Exact best-response evaluation exposed systematic failures hidden by sampled matches, then separated the effects of opponent diversity, adversarial backups, and symmetry-aware evidence pooling.
Read note