Duke NLP researcher focused on evaluating LLMs in health, from output-variability audits to search-agent grounding
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
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—
Not yet scored: no publication counts, venue records, or yearly output totals exist in this snapshot. This axis would measure overall research productivity — the volume and venues of published work — once output data is collected.
Impact—
Not yet scored: citation counts, h-index, or downstream-adoption measures are absent from the available data. This axis would quantify how widely the researcher's work is cited and built upon across the literature once impact metrics are collected.
Focus100/100
Scores 100/100 on maximum concentration: the field distribution (fieldDist) contains exactly one field — 'natural language processing (LLM evaluation in health)' — with no secondary areas in the record. Both 2026 catalog entries fit this lane: an audit of ChatGPT-versus-API output variability for health queries, and a factual-evidence attack on search agents, keeping the profile tightly scoped to evaluation-oriented NLP.
Collaboration51/100
Scores 51/100, derived from 12 distinct coauthors in the record. The frequent-collaborator list spans 12 named researchers — from senior Duke colleagues Bhuwan Dhingra and Monica Agrawal to junior collaborators such as Yulin Zhang, Yukun Huang, and Sanxing Chen. Both cataloged 2026 papers list this researcher in a middle author position without corresponding status, consistent with steady mid-sized team science rather than a large network or solo work.
Activity—
Not yet scored: beyond two 2026 arXiv entries, no longitudinal publication series or yearly activity counts are available. This axis would measure publishing cadence and sustained research engagement over time once longitudinal data is gathered.
Reproducibility55/100
Scores 55/100, computed from a single evaluated paper (reproPapers = 1, reproMean = 55) and therefore preliminary. The scored study is the 2026 health audit, whose Evolve review credits a sound paired ChatGPT-versus-API design but flags uncorrected p-values and LLM-judged features; its reproducibility readiness is rated medium, leaving clear headroom for stronger statistical reporting.
Research summary
Xunjian Yin is a researcher at Duke University working in natural language processing, with a cataloged concentration on evaluating large language models in health settings. The bibliometric field distribution in the available record contains exactly one field — 'natural language processing (LLM evaluation in health)' — which yields a maximum focus score of 100 and marks this as a tightly scoped, early-stage profile. Two 2026 arXiv papers anchor the Evolve catalog. The first, 'Challenges of Auditing: Variability in Outputs of Large Language Models for Health,' asks whether ChatGPT's chatbot interface and its API give systematically different answers to health queries; Evolve's review finds the paired audit design sound while cautioning that uncorrected p-values and LLM-judged features temper the claims, and rates reproducibility readiness as medium. The second, 'Lazy Grounding: Attacking Search Agents with Factual Evidence,' extends the evaluation lens toward search agents. The collaboration score of 51/100 rests on 12 distinct coauthors, including Duke colleagues Bhuwan Dhingra and Monica Agrawal, with this researcher in a middle author position on both papers. Output, impact, and activity dimensions remain unscored because the underlying bibliometric data has not yet been collected. The profile should be read as a name-resolved snapshot from public bibliographic data rather than a comprehensive career record.
Collaboration pattern
The collaboration score (51/100) is derived from 12 distinct coauthors in the record. The frequent-collaborator list is anchored at Duke, including senior colleagues Bhuwan Dhingra and Monica Agrawal alongside Yulin Zhang, Yukun Huang, Sanxing Chen, Tianyi Lin, Ziang Yang, Yewon Chang, and others. In both cataloged 2026 papers this researcher appears in a middle author position without corresponding-author status, consistent with a contributor role in small-to-mid-sized team projects rather than solo work or a large consortium.
Research keywords
Highlights
Systematic ChatGPT-versus-API output differences for health queries
Lazy Grounding: factual-evidence attacks on search agents
The second 2026 arXiv paper, 'Lazy Grounding: Attacking Search Agents with Factual Evidence,' extends the evaluation agenda from chatbot outputs to search agents, probing how agents handle adversarially supplied factual evidence. This paper has not yet been reviewed by Evolve, so no verdict, keywords, or reproducibility rating are available for it.
Auditability framed as a measurable property of deployed LLMs
Timeline
- 2026Published 'Challenges of Auditing: Variability in Outputs of Large Language Models for Health' on arXiv, in a middle author position.
- 2026Published 'Lazy Grounding: Attacking Search Agents with Factual Evidence' on arXiv, also in a middle author position.
- 2026The health-audit paper entered the Evolve review catalog with a 'medium' reproducibility readiness rating and a verdict crediting its sound paired design.
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 dimensions are computed from public bibliometric data (OpenAlex-sourced platform data); coverage of the publication record may be incomplete.
- The reproducibility score rests on a single evaluated paper (reproPapers = 1, reproMean = 55) and should be read as preliminary.
- 'Lazy Grounding: Attacking Search Agents with Factual Evidence' has not yet been reviewed by Evolve — no verdict, field, keywords, or reproducibility rating exist for it.
- This profile is name-resolved from public bibliographic data anchored to ORCID 0000-0002-2742-2225; no biography, award, or career-history information is included.