Shirley Wu
Shirley Wu is an AI researcher who works on post-training at Meta TBD Labs. She earned a data science B.S. at the University of Science and Technology of China and began a computer science PhD at Stanford in 2022. She defended her thesis, Collaborative AI Agents, in January 2026. Her research includes Med-Flamingo for medical question answering, STaRK for LLM retrieval, and CollabLLM for multi-turn interaction; CollabLLM received ICML's 2025 Outstanding Paper Award. In 2026, she led HumanLM, a user simulator built around latent-state alignment, and coauthored Optimas, a framework for optimizing compound AI systems.
Timeline
-
2018–2022
A data science degree
She earns a data science B.S. at the University of Science and Technology of China in 2022 and completes undergraduate research at the National University of Singapore from 2020 to 2021.
-
2022
A PhD at Stanford
She begins a computer science PhD at Stanford in 2022 and studies LLM agents, alignment, and human-agent interactions.
-
2023
Med-Flamingo for medical AI
In 2023, she coauthors Med-Flamingo, a multimodal few-shot model for medical visual question answering.
-
2024
STaRK benchmarks LLM retrieval
In 2024, she coauthors STaRK, a retrieval benchmark for product search, academic papers, and precision medicine.
-
2025
CollabLLM wins an ICML award
In 2025, she coauthors CollabLLM, which trains language models through simulated multi-turn interactions. ICML gives it an Outstanding Paper Award.
-
January 2026
Leading HumanLM
She defends her Stanford PhD thesis, Collaborative AI Agents, in January 2026 and releases HumanLM, her final leading PhD project. The model trains user simulators to align latent states with responses through reinforcement learning.
-
2026
Optimizing compound AI systems
In 2026, she coauthors Optimas, an ICLR paper that uses local rewards aligned with system performance to optimize compound AI systems.
-
As of October 2026
Post-training at Meta
As of October 2026, she works on post-training at Meta TBD Labs.
Sources and credits
Photo: We have no published photo of Shirley Wu yet. Send us a free-licensed photo.
- Shirley Wu's Homepage
- STaRK: Benchmarking LLM Retrieval on Textual and Relational Knowledge Bases
- CollabLLM: Teaching LLMs to collaborate with users
- ICML 2025 Awards
- HumanLM: Simulating Users with State Alignment Beats Response Imitation
- Optimas: Optimizing Compound AI Systems with Globally Aligned Local Rewards