UW NLP researcher: long-context reasoning, distillation during LLM mid-training, and reproducible evaluation
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
5
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
An output score is not yet available: OpenAlex coverage is too incomplete for a reliable productivity score. This axis would measure publication volume and yearly consistency; the partial signal — 39 works over the last three years — is not yet fully attributed.
Impact—
Impact scoring is unavailable: citation coverage in OpenAlex is too thin for this profile. The axis would measure citation-based influence (h-index, i10-index); current raw signals — h-index 5, i10-index 2 — are reported but not yet scored.
Focus100/100
Focus scores 100/100: fieldDist assigns the reviewed catalog entirely to natural language processing, and topic mass concentrates in topic modeling (25) and NLP techniques (17), with multimodal applications (12) a secondary thread. Both 2026 reviewed papers sit squarely in LLM research — long-context reasoning and mid-training distillation — confirming a tightly concentrated, single-field agenda.
Collaboration56/100
Collaboration scores 56/100 with 16 distinct coauthors and 12 recurring partners, indicating a moderate-breadth network rather than a large or solitary one. Recurring names include Jacqueline He, Howard Yen, Margaret Li, Zhuokai Zhao, and senior researchers Pang Wei Koh, Luke Zettlemoyer, and Wen-tau Yih; both reviewed papers are middle-author, non-corresponding positions, consistent with contribution inside larger teams.
Activity100/100
Activity scores 100/100: yearly works rise from 6 (2022) and 4 (2023) to 7 (2024), 14 (2025), and 18 (2026) — 39 works in the last three years, with 2026 alone contributing 18. The trajectory shows steep, sustained acceleration rather than sporadic bursts, and both cataloged 2026 arXiv papers reinforce the picture of an exceptionally active current output pace.
Reproducibility100/100
Reproducibility scores 100/100 (reproMean 100 across 1 evaluated paper). Evolve rates both reviewed papers 'high' reproducibility readiness: the R3CON context-representation harness ships solid ablations against 9 baselines, and the distillation study rests on controlled OLMo-2 experiments at 1B scale. The small evaluated base (1 paper) means the perfect score is promising early evidence, not a settled track record.
Research summary
Shuyue Stella Li is a natural language processing researcher at the University of Washington whose recent work concentrates on how large language models represent, use, and are trained to reason with context. OpenAlex bibliometrics show a steeply accelerating record: yearly works rising from 6 (2022) and 4 (2023) to 7 (2024), 14 (2025), and 18 (2026), with 39 works in the last three years, an h-index of 5 and i10-index of 2, and 16 distinct coauthors. Two 2026 arXiv papers in the Evolve catalog illustrate the agenda. 'Realize What Matters' introduces R3CON, a principled corpus-representation harness for large-scale reasoning that improves over 9 baselines by +19.9pp on CorpusQA and +8.4pp on Loong, with solid ablations. 'Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall' uses controlled OLMo-2 experiments to show mid-training distillation trades factual recall for reasoning, and that an entropy-routed Switch Distillation largely fixes the trade-off at 1B scale. Topic distribution spans topic modeling (25), NLP techniques (17), multimodal applications (12), speech and dialogue (6), and recommender systems (5), with a single primary field (natural language processing). Evolve rates both reviewed papers 'high' for reproducibility readiness (reproMean 100, 1 evaluated paper). Output and impact dimensions are not yet scored.
Collaboration pattern
With 16 distinct coauthors and 12 recurring partners, Li works in a mid-sized collaborative network. Frequent coauthors include Jacqueline He, Howard Yen, Margaret Li, Zhuokai Zhao, and Qiang Zhang, alongside senior researchers Pang Wei Koh, Luke Zettlemoyer, Wen-tau Yih, and Yinglong Xia. Both reviewed 2026 papers place Li in a middle, non-corresponding author position, suggesting contributions within larger research teams rather than solo-led projects. The 56/100 collaboration score reflects this moderate breadth.
Research keywords
Highlights
R3CON: principled context representation for large-scale reasoning
Mid-training distillation trades factual recall for reasoning
Steep output acceleration in 2025–2026
Yearly works climbed from 7 (2024) to 14 (2025) and 18 (2026) — 39 works over the last three years, the basis for a 100/100 activity score.
Timeline
- 20226 works recorded in OpenAlex — the earliest year in the bibliometric window.
- 20234 works recorded — the lowest annual count in the window.
- 20247 works; the start of a sustained upward climb.
- 202514 works — output doubles year over year.
- 202618 works recorded, including both Evolve-cataloged arXiv papers (R3CON and mid-training distillation); 39 works across 2024–2026.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric statistics are computed from OpenAlex only (identity: openalex); no ORCID link was available, so the profile is name-resolved from public bibliographic data and name collisions cannot be fully excluded.
- Only two papers have been reviewed by Evolve; the reproducibility score rests on a single evaluated paper (reproPapers 1).
- Output and impact dimensions are unscored due to insufficient underlying data; h-index (5) and i10-index (2) may undercount recent 2025–2026 work due to citation lag.
- Topic and field labels come from automated OpenAlex classification and may misassign individual works.