UC Berkeley NLP researcher in LLM-based human simulation and gaze-guided vision-language pragmatics
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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—
No output metrics are available in this snapshot, so the axis is unscored. It would measure publication volume and consistency over time; the current evidence base is just two 2026 arXiv papers in the Evolve catalog.
Impact—
Impact data such as citations or downstream adoption is not yet available, so this axis is unscored. It would measure how widely the 2026 catalog papers are cited and built upon — evidence that has had little time to accumulate.
Focus0/100
The focus score is 0 on the 0-100 scale, derived from a field distribution split exactly 0.5/0.5 between two NLP subfields — LLM-based human simulation and vision-language pragmatics — with no dominant line. The two 2026 catalog papers sit one in each thread, indicating low concentration and a deliberately two-track early research agenda.
Collaboration44/100
The collaboration score is 44, based on 8 distinct coauthors across the two catalog papers. The frequent-coauthor list — Alane Suhr, Minwoo Kang, Seun Eisape, Ayush Raj, Suhong Moon, Joseph Suh, David M. Chan, and John Canny — indicates compact, Berkeley-centered teams. Wright is first author on the gaze paper and a middle author on the human-simulation paper, with overlapping collaborator groups spanning both lines.
Activity—
Recent-activity signals such as publication cadence and preprint frequency are insufficient in this snapshot, so the axis is unscored. It would measure how actively the researcher posts and updates work; the current sample spans a single year, 2026.
Reproducibility78/100
The reproducibility score is 78, computed as a mean of 78 across 2 reviewed papers. Evolve rates the first-author gaze paper high and the persona-mixture paper medium; the verdicts credit strong ablations and human evaluation while flagging missing significance tests and unexamined data contamination as the main gaps.
Research summary
Téa Wright is a researcher at the University of California, Berkeley working in natural language processing. This is an early-stage profile built from two 2026 arXiv papers in the Evolve catalog. The field distribution splits evenly — 0.5 and 0.5 — between LLM-based human simulation and vision-language pragmatics. As first author of "Learning to Refer from Estimated Listener Gaze," Wright shows that reward signals from estimated listener gaze scanpaths make vision-language speakers more pragmatic in reference games; Evolve's review credits strong ablations and human evaluation with high reproducibility readiness, while noting no significance tests and weak cross-backbone generalization. As a middle author on "Pretrained Persona Mixture Models and Tandem Models for Human Simulation," Wright contributes to evidence that pretrained persona mixture base models beat instruction-tuned LLMs at human simulation, with tandem selection performing best, though data contamination is unexamined and readiness is medium. The reproducibility mean across the two reviewed papers is 78, with 8 distinct coauthors including Alane Suhr and John Canny. Output, impact, and activity metrics are not yet available.
Collaboration pattern
Wright's two catalog papers (both 2026) share a pool of 8 distinct coauthors, yielding a collaboration score of 44. The frequent-coauthor list — Alane Suhr, Minwoo Kang, Seun Eisape, Ayush Raj, Suhong Moon, Joseph Suh, David M. Chan, and John Canny — points to compact, Berkeley-anchored teams. Wright first-authored the gaze paper and contributed in a middle position to the human-simulation paper, with overlapping collaborator groups across the two lines.
Research keywords
Highlights
Gaze-derived rewards make VLM speakers more pragmatic
Persona mixture base models beat instruction-tuned LLMs at human simulation
Early two-track NLP agenda with a solid reproducibility signal
Across the two reviewed 2026 papers, reproducibility readiness is high once and medium once, averaging 78, while the field split is an even 0.5/0.5 between LLM-based human simulation and vision-language pragmatics — a deliberately two-track early agenda grounded in evaluated, mostly reproducible work.
Timeline
- 2026First-authored "Learning to Refer from Estimated Listener Gaze" (arXiv) in vision-language pragmatics; Evolve rates its reproducibility readiness high.
- 2026Middle-authored "Pretrained Persona Mixture Models and Tandem Models for Human Simulation" (arXiv) in LLM-based human simulation; readiness rated medium.
- 2026Both papers enter the Evolve catalog the same year, splitting 0.5/0.5 across two NLP subfields with 8 distinct coauthors and a reproducibility mean of 78.
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.
- Dimension scores derive from OpenAlex-based bibliometrics; only two 2026 arXiv papers are in the Evolve catalog, so output, impact, and activity remain unscored.
- The reproducibility score of 78 averages just 2 reviewed papers — a small sample.
- No ORCID or country is provided; "Téa Wright" is name-resolved from public bibliographic data and identity should be verified before reuse.