ML researcher at Shandong University probing model collapse and data-efficient training via rigorous benchmarks
3
papers here
3
reviewed by Evolve
3
first-author
0
corresponding
—
works · OpenAlex
1
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 output score is computed yet: the OpenAlex record lacks complete publication aggregates for this profile, so total volume and venue-weighted productivity cannot be fairly estimated. This axis would measure cumulative works, publication venues, and year-over-year output throughput.
Impact—
No impact score is available: beyond h-index 1 and i10-index 1, citation aggregates are absent, and the three catalog preprints are 2026 works too recent to accrue citations. This axis would measure citation influence and community uptake of the research.
Focus0/100
The focus score is 0/100, and the raw topics explain why: five OpenAlex topics — topic modeling, NLP techniques, software testing and debugging, adversarial robustness, and ML algorithms — are tied at two mentions each, while the field distribution splits evenly at one-third across three ML subfields. The agenda is broad and thematically dispersed rather than concentrated, a typical early-stage exploration pattern.
Collaboration14/100
Collaboration scores 14/100, consistent with the raw counts: OpenAlex lists exactly one distinct coauthor, and the frequent-coauthor list names Zhongyi Han alone. All three catalog papers are first-authored, with no corresponding-author role recorded. This is a tightly paired, early-stage pattern — deep joint work with one partner rather than a wide network — so cross-group data and compute access may be the main scaling constraint.
Activity61/100
Activity scores 61/100 on a sharply front-loaded record: OpenAlex countsByYear show one work in 2019 and seven in 2026, with recent3yWorks at 7 — effectively the entire visible output lands in the current year. The three 2026 first-author preprints in the Evolve catalog drive this burst, indicating a productive but very recent phase whose sustained pace cannot yet be tested across multiple years.
Reproducibility85/100
Reproducibility is the strongest measured axis at 85/100, computed as reproMean across three assessed papers (reproPapers: 3). Two catalog papers carry high repro-readiness — the coreset benchmark and the oligopoly null result — while the recursive-training fragility study is rated medium. Evolve's verdicts repeatedly cite careful, audited designs, with caveats about partly single-seed ImageNet duels and small 1–4B corpus regimes rather than methodological weakness.
Research summary
Yangze Liu is a machine learning researcher at Shandong University whose recent work examines when and how language-model training degrades — and whether the data curation methods meant to help are actually worth their cost. The Evolve catalog lists three first-author arXiv preprints from 2026: an audited benchmark of coreset selection under matched wall-clock budgets, a 13-model study mapping checkpoint-specific collapse fragility in recursive training, and a null result showing market concentration barely steers model collapse in 1–4B multi-model ecosystems. OpenAlex records a small but sharply recent footprint: an h-index and i10-index of 1, one work in 2019, and seven in 2026, with a single distinct coauthor — Zhongyi Han — indicating a concentrated, early-stage collaboration pattern. Topic assignments span topic modeling, NLP techniques, software testing and debugging, adversarial robustness, and ML algorithms, each tied at two mentions, while the field distribution splits evenly across three machine-learning subfields. The strongest measurable signal is reproducibility: three assessed papers average 85/100, with two rated high readiness. The profile is name-resolved from public bibliographic data.
Collaboration pattern
The visible collaboration footprint is minimal: OpenAlex records one distinct coauthor, and Zhongyi Han is the sole frequent collaborator, consistent with the low 14/100 collaboration score. All three Evolve-catalog papers are first-authored, with no corresponding-author role recorded. The pattern suggests a stable two-person working unit producing first-author arXiv preprints; the data shows no evidence of larger consortia, industry partnerships, or multi-institution projects.
Research keywords
Highlights
Coreset selection fails its cost-benefit test under matched budgets
Mapping a fragility spectrum for recursive language-model training
A careful null: market concentration barely steers model collapse
Timeline
- 2019Earliest OpenAlex-indexed output appears, with exactly one work recorded in that year's count.
- 2026Output jumps to seven indexed works — matching the recent3yWorks total of 7 — marking a sharp activity burst.
- 2026First-author arXiv preprint 'Are Coreset Selection Methods Worth Their Cost?' is cataloged by Evolve with high repro-readiness.
- 2026First-author arXiv preprint 'A Fragility Spectrum for Recursive Language-Model Training' is cataloged, based on a 13-model study.
- 2026First-author arXiv preprint 'The Oligarch Barely Steers Model Collapse in Multi-Model Ecosystems' is cataloged as a high-reproducibility null result.
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.
- Bibliometrics are OpenAlex-derived (identity: openalex); no ORCID is linked, so the profile is name-resolved from public bibliographic data and common-name ambiguity cannot be fully excluded.
- CountsByYear is sparse — only 2019 (1 work) and 2026 (7 works) are present — so mid-period output is not observable.
- Output and impact dimensions are unscored (dimMissing); h-index 1 and i10-index 1 are the only citation-side figures provided.
- The Evolve catalog covers three 2026 arXiv preprints; peer-reviewed venues, awards, and funding information fall outside the data scope.