Shuyue Stella Li

Shuyue Stella Li is a natural language processing researcher who studies how language models seek information, handle uncertainty and adapt to people. She earned B.S. and M.S.E. degrees at Johns Hopkins, then researched LLVM optimization and cross-lingual speech features. She coauthored MediQ for clinical reasoning, ValueScope for social norms and multilingual reliability, ALFA for question asking, Spurious Rewards in reinforcement learning and the open Olmo 3 model family. In 2026 she coauthored PrefDisco; as of October 2026, she is a University of Washington PhD student and visiting researcher at Meta FAIR.

Now at: FAIR (Meta)
Works: 55
Citations: 266
h-index: 5

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2022–2023

Shackleton and cross-lingual speech

Before graduate school, she earns B.S. and M.S.E. degrees at Johns Hopkins, studying cognitive science, computer science and applied mathematics. In 2022, she coauthors a genetic-improvement study that uses Shackleton to search for optimized LLVM pass sequences. In 2023, she starts a PhD at the University of Washington and coauthors a study of cross-lingual speech features that introduces the Phonetic-Syntax Ratio.
2024

MediQ for clinical reasoning

In 2024, she coauthors MediQ, an interactive benchmark and question-asking system for clinical reasoning. It simulates patient and expert exchanges with incomplete information and reports a 22.3% diagnostic-accuracy improvement from confidence-based abstention.
2024

ValueScope and multilingual abstention

In 2024, she coauthors ValueScope, a framework that uses language models to measure social norms across 13 Reddit communities. Also in 2024, she coauthors work that teaches multilingual LLMs to abstain by learning from feedback across languages, improving results for low-resource languages by up to 9.2%.
2025

ALFA for clinical question asking

In 2025, she coauthors ALFA, which trains language models to ask clearer, more relevant questions by breaking question quality into attributes and using preference-based optimization.
2025

Spurious rewards in RLVR

In 2025, she coauthors Spurious Rewards, which finds that RLVR improves Qwen2.5-Math-7B's MATH-500 performance by 21.4 percentage points with random rewards, close to the 29.1-point gain from ground-truth rewards.
2025

Olmo 3's open model flow

In 2025, she coauthors Olmo 3, a family of fully open language models at 7B and 32B parameter scales. The release includes the full model flow, with stages, checkpoints, data and dependencies.
2026

PrefDisco for personalized reasoning

In 2026, she coauthors PrefDisco, a benchmark for proactive personalized reasoning where models ask users about their preferences and adapt their reasoning and responses.
As of October 2026

UW PhD and Meta FAIR research

She starts a PhD at the University of Washington in 2023, advised by Yulia Tsvetkov. As of October 2026, she is a PhD student at UW and a visiting researcher at Meta FAIR, working with Asli Celikyilmaz and Luke Zettlemoyer.