About
Hi - I'm Marianne Menglin Liu, a Principal Applied Scientist on the Generative AI Science team at Oracle AI (OCI), where I work on natural language interfaces for enterprise AI and data platforms. I'm based in the San Francisco Bay Area.
I focus on industrial AI research, building production-grade systems that connect large language models with enterprise data and tools. My work centers on NL2SQL, agentic reasoning, and data-centric AI, with an emphasis on evaluation, reliability, and real-world deployment.
At Oracle, I have led enterprise NL2SQL and agentic data systems, LLM evaluation frameworks, multimodal synthetic data platforms, and anomaly detection services on OCI. In NL2SQL, I contributed to #1 results on Archer NL2SQL and Spider 2.0 Lite. Previously, I was a Senior Data Scientist at IBM, where I led clinical-driven AI systems from design to production for healthcare and pharmaceutical organizations.
I hold a Master's in Information Systems from Carnegie Mellon University and a Bachelor of Management (Honors) from Fudan University; together with my earlier experience in management and technical consulting, this interdisciplinary social science background shaped my interest in interdisciplinary, real-world AI research.
News
PLSQLBench: Benchmarking LLM Systems for Executable Procedural Database Programming has been accepted to the EMNLP 2026 Industry Track.
My paper, Routesplain: Towards Faithful and Intervenable Routing for Software-Related Tasks, was accepted to the 3rd Conference on Language Modeling (COLM 2026).
We released the SOMA-SQL paper on arXiv. SOMA-SQL reached #1 on the Spider 2.0 Lite leaderboard with 72.02% accuracy.
My papers ToolScope and LAD-RAG were accepted to the ACL Main Conference 2026.
#1 overall on the bilingual leaderboard in the 2025 Archer NL2SQL Challenge; I represented Oracle and collaborated with Ashish Somayajula to deliver +6.3 points in English and +12.6 points in Chinese over the runner-up, while maintaining over 99% SQL validity.