Yunshi Lan

Applied Decision Science (United States) · China

Unclaimed · AI-structured profileORCID OpenAlex

NLP and LLM researcher spanning topic modeling, on-policy distillation, value alignment, and multimodal AI education

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) 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 publication-volume metric is available in this snapshot, so output is unscored. This axis would measure total indexed works and year-by-year production volume; adding venue-level counts would resolve it.

Impact—

An h-index of 17 and i10-index of 22 are recorded, but the citation detail behind an impact score (total citations, citing works, field-normalized influence) is absent, so this axis stays unscored.

Focus0/100

Focus scores 0 because the OpenAlex field distribution is perfectly flat: AI in education (multimodal generation), LLM post-training and distillation, and LLM value alignment and safety evaluation each hold exactly 0.333. Topic counts reinforce the breadth — topic modeling (70) and NLP techniques (55) lead, with multimodal ML (28), semantic web and ontologies (12), and speech and dialogue systems (10) trailing. This is a deliberately broad portfolio rather than a single specialized line.

Collaboration55/100

Collaboration scores 55 from OpenAlex co-authorship: 15 distinct co-authors across the recent three-year window covering 48 works — a moderate network size. In Evolve's catalog, Yunshi Lan sits in a middle author position on all three 2026 papers, never as corresponding author, indicating a contributor role within larger teams. Recurring partners include Xinchen Ma, Shuimu Wang, Gaole He, Yanbin Zhang, Chunyang Wang, and Weining Qian, spanning the distillation, education, and alignment projects.

Activity100/100

Activity scores 100, the maximum. OpenAlex counts by year show output escalating from 5 works in 2022 to 20 in 2023, 19 in 2024, 12 in 2025, and 17 already indexed in 2026, with 48 works across the last three years. That cadence marks a sustained, high-intensity production phase rather than episodic bursts.

Reproducibility58/100

Reproducibility scores 58, the mean readiness across 3 Evolve-reviewed papers: high for the Direct-OPD selective-supervision distillation study (limits are single-run, math-only evaluation), medium for the value-conflict probing paper (judge dependence and a choice-format confound), and low for the PIVOT video-tutor work (evaluation lacking statistical rigor). A genuine but polarized replication signal.

Research summary

Yunshi Lan is an OpenAlex-identified researcher whose record spans topic modeling, natural language processing, and multimodal machine learning — the three largest topic clusters in this profile, with 70, 55, and 28 indexed works respectively, ahead of semantic web and ontologies (12) and speech and dialogue systems (10). Current work concentrates on three evenly weighted directions: artificial intelligence in education (multimodal generation), LLM post-training and distillation, and LLM value alignment and safety evaluation, each holding exactly one third of the recent field distribution. The productivity curve has steepened sharply: from 5 indexed works in 2022 to 20 in 2023, 19 in 2024, 12 in 2025, and 17 already indexed in 2026, giving 48 works in the last three years, an h-index of 17 and an i10-index of 22. Evolve has reviewed three 2026 arXiv papers — selective token-level supervision for on-policy distillation, pedagogy-guided generative video tutors for STEM learning, and value-conflict probing of LLM value expression — all with middle-author placement and none as corresponding author. Reproducibility readiness across them is mixed (high, medium, low), averaging 58. With 15 distinct recent co-authors and a collaboration score of 55, this reads as a broadly engaged, highly active contributor working across several fast-moving LLM research fronts.

Collaboration pattern

Co-authorship is moderate: OpenAlex records 15 distinct co-authors across 48 works in the last three years, giving a collaboration score of 55. In the Evolve catalog, Yunshi Lan appears as a middle author on all three 2026 papers and is never the corresponding author, pointing to a contributor role within larger teams. Recurring partners include Xinchen Ma, Shuimu Wang, Gaole He, Yanbin Zhang, Chunyang Wang, Weining Qian, and Yibo Zhao, spanning the distillation, education, and value-alignment projects.

Research keywords

Topic modelingNatural language processingMultimodal machine learningLLM post-trainingOn-policy distillationWeak-to-strong transferToken-level rewardsLLM value alignmentValue conflicts and safety evaluationEducational video generationIntelligent tutoring systemsSpeech and dialogue systems

Highlights

Selective token-level supervision improves on-policy distillation
Evolve's review finds Direct-OPD's log-ratio reward ignores probability mass, while masking low-JSD states wins in 7 of 8 settings — sound theory and honest evaluation, though limited to single runs and math-only tasks. Reproducibility readiness is rated high.
PIVOT: pedagogy-guided generative video tutors for STEM learning
The paper adds pedagogy, verification, and misconception-aware remediation into generated STEM videos, shifting video generation from content creation to learning support. Evolve calls it promising but notes the evaluation lacks statistical rigor and the human studies carry confounds; reproducibility readiness is low.
Value-conflict probing maps the dynamics of LLM value expression
Using Schwartz's theory of basic values, the study finds LLM value expression to be dynamic yet bounded when values conflict. Evolve credits its strong validation but tempers the conclusions because of LLM-as-a-judge dependence and a choice-format confound under pressure; reproducibility readiness is medium.

Timeline

2016Earliest year in the OpenAlex counts-by-year record, with 1 indexed work.
2023Output surged to 20 indexed works, up from 5 in 2022 — the start of the current high-productivity phase.
2024Sustained the pace with 19 indexed works, confirming the previous year's jump was not a one-off.
2026Three Evolve-reviewed arXiv papers appeared across on-policy distillation, pedagogy-guided video tutors, and LLM value-conflict probing; 17 works are already indexed this year.

Papers in the Evolve catalog

Frequent co-authors

Xinchen Ma×1 Shuimu Wang×1 Gaole He×1 Yanbin Zhang×1 Chunyang Wang×1 Yunshi Lan×1 Weining Qian×1 Yibo Zhao×1 Zixuan Yang×1 Xiang Li×1 Kaicheng Zhang×1 Jingyi Xiao×1

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 indicators (h-index, yearly counts, topics, co-authors) come from an OpenAlex-derived snapshot; citation-level impact detail is not included.
Affiliation metadata is inconsistent — the record lists 'Applied Decision Science (United States)' while the country field says 'CN' — so the affiliation should be treated as unresolved.
The profile is name-resolved from public bibliographic data (OpenAlex identity, ORCID 0000-0002-0192-8498 as given) and may merge or split same-name authors.
Quality judgments rest on only three Evolve-reviewed papers, all 2026 arXiv preprints; peer-review status and venue acceptance are unknown.