Shandong University ML researcher studying model collapse, synthetic data, and benchmark rigor in LLM training
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
Output measures publication volume and consistency — total works, per-year counts, and authorship position — but this profile's record is incomplete (only 4 works appear, all dated 2026), so no score is reported rather than estimated.
Impact—
Impact would measure citation-based influence — h-index, i10-index, and citations per paper. The snapshot shows h-index 0 and i10-index 0 with no reliable citation stream yet, so this axis is left unscored instead of guessed.
Focus0/100
Focus scores 0/100, matching the raw field distribution: the three catalogued fields each take exactly one third (LLM training dynamics/model collapse; machine learning/NLP; data-efficient training). The directions are closely adjacent around recursive training and synthetic data, but no single field yet dominates, so measured concentration is zero.
Collaboration14/100
Collaboration scores 14/100. The OpenAlex snapshot records just 1 distinct co-author, and Yangze Liu is the sole frequent collaborator, appearing across all three catalogued 2026 preprints with Zhongyi Han in last position. This points to a compact two-person team rather than a broad co-authorship network.
Activity47/100
Activity scores 47/100. All 4 works in the snapshot are dated 2026 (countsByYear shows 2026: 4; recent3yWorks = 4), so the record reflects a recent, concentrated burst of arXiv preprints rather than sustained multi-year publishing — consistent with an early-stage output profile.
Reproducibility85/100
Reproducibility is the strongest axis at 85/100, computed over 3 papers (reproMean 85). Evolve rates two preprints high readiness — the coreset benchmark and the oligopoly study — and one medium, the fragility-spectrum paper, reflecting an audit-driven, benchmark-heavy empirical style.
Research summary
Zhongyi Han is a machine learning researcher at Shandong University whose profile is built from public bibliographic data (OpenAlex) and three arXiv preprints reviewed by Evolve, all from 2026 and all with Han in the last author position. The work concentrates on the training dynamics of large language models under synthetic data: one preprint charts a "fragility spectrum" of recursive language-model training, finding checkpoint-specific model-collapse fragility (a 5× u4 spread with seed-stable ordering) across a 13-model study in a 1–4B-parameter corpus regime; a companion study reports a careful null result — in multi-model ecosystems, market concentration barely steers the pace or endpoint of model collapse. A third preprint asks whether coreset selection methods are worth their wall-clock cost, and an audited benchmark finds no selector wins under matched budgets. This set gives the profile a distinctive empirical, audit-and-benchmark orientation: the reproducibility axis is the strongest at 85/100 across all 3 reviewed papers, while activity is moderate (47/100, with all 4 recorded works in 2026) and the co-authorship footprint is thin (14/100; 1 distinct co-author, with Yangze Liu the recurring collaborator). Output and impact are not scored because the citation record is not yet available; h-index and i10-index read 0 in the current snapshot. Focus is currently low (0/100) as effort splits evenly across three adjacent directions.
Collaboration pattern
Co-authorship is narrow: the OpenAlex snapshot records 1 distinct co-author, and Yangze Liu is the sole frequent collaborator, appearing on all three catalogued 2026 preprints with Zhongyi Han in last author position (none marked corresponding). The pattern suggests a compact two-person team producing benchmark and empirical studies rather than a distributed multi-lab network; no co-authorship clusters beyond Shandong University are visible in the provided data.
Research keywords
Highlights
Coreset selection under matched wall-clock budgets: no selector wins
A fragility spectrum for recursive language-model training
Market concentration barely steers model collapse
Timeline
- 2026Posts "Are Coreset Selection Methods Worth Their Cost?", an audited benchmark of data-efficient training under matched wall-clock budgets.
- 2026Posts "A Fragility Spectrum for Recursive Language-Model Training", mapping checkpoint-specific collapse fragility across 13 models.
- 2026Posts "The Oligarch Barely Steers Model Collapse in Multi-Model Ecosystems", a null result on market concentration in 1–4B ecosystems.
- 2026All 4 works recorded in the OpenAlex snapshot fall in 2026, with Zhongyi Han in last author position on each catalogued paper.
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 computed from an OpenAlex snapshot only; h-index and i10-index read 0 and no citation-impact data is available.
- Coverage is limited to the three arXiv preprints reviewed by Evolve; peer-reviewed venues, DOIs, and an ORCID are not recorded.
- "Zhongyi Han" is a common name; this profile is name-resolved from public bibliographic data, so authorship attribution may be imperfect.
- Radar scores derive from the same limited snapshot and may shift materially as the publication record matures.