Dahua Lin

Chinese University of Hong Kong · Hong Kong SAR China

Unclaimed · AI-structured profileORCID OpenAlex

CUHK ML researcher spanning LLM pretraining, chemical reaction data infrastructure, and scientific document understanding.

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—

Not computed: this snapshot lacks the work-level completeness needed to score raw publication volume reliably. The output axis would measure publication volume, productivity trends, and venue spread across the career record.

Impact—

Not computed: citation-linked evidence is insufficient in the current snapshot to derive a verified impact score. The impact axis would measure citation density, citing works, and downstream influence; raw provider indices exist but remain unscored here.

Focus0/100

Scores 0/100, reflecting an unusually broad spread: the field distribution splits almost evenly across machine learning (LLMs), cheminformatics/AI for chemistry, and multimodal scientific document understanding, while top OpenAlex topics — neural network applications (160), multimodal ML (121), GANs and image synthesis (97), topic modelling (92), domain adaptation (75) — span distinct subfields rather than one concentrated line.

Collaboration85/100

Scores 85, grounded in 71 distinct co-authors over the trailing three years and team-scale authorship on the reviewed papers — the Intern-NCP team appears among frequent collaborators, and Lin is last author on SciDocBench. Recurring named partners include Bowen Zhou, Jiaqi Cao, Chiyu Chen, Shuang Cheng, and Xu Cheng, pointing to sustained group-driven projects.

Activity100/100

Scores 100: 208 works in the trailing three years, with annual counts rising from 9 (2005) through 14 (2017), 27 (2018), and 38 (2020) to a record peak of 115 (2024), followed by 84 (2025) and 9 recorded so far in 2026 — a steep, sustained acceleration in publication tempo.

Reproducibility58/100

Scores 58 (reproMean) across 3 Evolve-reviewed papers, and readiness varies widely: NCP-ArchPreview rates high; SciDocBench medium, with a sound protocol but small N and no confidence intervals; DianShi-RxnDB low, with pipeline details deferred and no downstream benchmark shown. Reproducibility thus hinges on artifact release for infrastructure-heavy work.

Research summary

Dahua Lin is a researcher at the Chinese University of Hong Kong whose recent work spans three near-equal strands: large language models, AI-for-chemistry reaction data infrastructure, and multimodal scientific document understanding, each taking roughly one third of the field distribution in the OpenAlex record. The bibliometric footprint is substantial: an h-index of 85, an i10-index of 234, and 208 works over the trailing three years, with annual counts climbing from 9 in 2005 to a peak of 115 in 2024 and 84 in 2025. The activity dimension scores 100 and collaboration 85, the latter reflecting 71 distinct co-authors in the recent window and consortium-style authorship such as the Intern-NCP team. Focus scores 0, consistent with the evenly split field distribution and top topics that range from neural network applications (160) to topic modelling (92). Evolve has reviewed three 2026 arXiv preprints: NCP-ArchPreview on latent-space language modelling via next concept prediction, DianShi-RxnDB claiming a 24M-reaction patent-mined database, and SciDocBench, a workflow-centred scientific document benchmark. Reproducibility averages 58 across these three, with readiness ratings from high to low. Output and impact dimension scores are not computed in this snapshot; the raw citation indices are reported as provided, without independent verification.

Collaboration pattern

Collaboration is extensive and team-based: 71 distinct co-authors in the trailing three years (dimension score 85). The reviewed 2026 preprints carry large consortium-style author lists — 'The Intern-NCP Team' appears among frequent collaborators, and Lin is last author on SciDocBench. Recurring named partners include Bowen Zhou, Jiaqi Cao, Chiyu Chen, Shuang Cheng, Xu Cheng, and Beiya Dai. The pattern suggests lab- and team-driven projects rather than small bilateral collaborations.

Research keywords

Large language modelsLatent-space language modelsAutoregressive pretrainingVector quantizationChemical reaction databasesPatent mining and information extractionAI for chemistryScientific document understandingMultimodal benchmarksEvidence groundingLLM-as-a-judgeDomain adaptation and few-shot learning

Highlights

Next Concept Prediction: toward latent-space language models
The NCP-ArchPreview technical report proposes joint token+concept prediction at 8.9B parameters, claiming 1.95× token-efficient pretraining. Evolve notes Stage-1 gains of +2.45 shrinking to +0.59 in Stage-2, with no significance tests; reproducibility readiness is rated high.
DianShi-RxnDB: a 24M-reaction database mined from patents
A fully automated pipeline builds a 24M-reaction organic chemistry database with a claimed 92.95% field accuracy, edging Pistachio. Evolve finds the scale credible but notes pipeline details are deferred and no downstream benchmark is shown; readiness is rated low.
SciDocBench: a workflow-centred scientific document benchmark
A workflow-centred benchmark for scientific document understanding (124 questions expanded to 496 instances) plus 15K SFT and 8K RL samples. Evolve rates the protocol sound but flags small N, no confidence intervals, and an unreported training evaluation in §5; readiness medium, with Lin as last author.

Timeline

2005Earliest year in the OpenAlex output record, with 9 works — the baseline before years of modest annual counts.
2018Output accelerates: annual works jump from 14 (2017) to 27, beginning a multi-year climb.
202038 works recorded, continuing the steady upward trajectory of the 2017–2021 stretch.
2024Peak year in the record: 115 works, up sharply from 53 in 2023; 84 follow in 2025.
2026Three Evolve-reviewed arXiv preprints appear — NCP-ArchPreview, DianShi-RxnDB, and SciDocBench — within 9 works recorded so far this year.

Papers in the Evolve catalog

Frequent co-authors

Bowen Zhou×2 The Intern-NCP Team×1 :×1 Jiaqi Cao×1 Chiyu Chen×1 Shuang Cheng×1 Xu Cheng×1 Beiya Dai×1 Yufan Feng×1 Kewen Ge×1 Ruijun Ge×1 Jiayi Huang×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.

Bibliometrics are derived from OpenAlex only (identity: openalex); citation counts and indices have not been independently audited by Evolve.
Output and impact dimension scores are missing in this snapshot; the h-index (85) and i10-index (234) are raw provider figures, not scored dimensions.
All three reviewed papers are 2026 arXiv preprints rather than peer-reviewed publications, and the reproducibility mean (58) rests on only 3 papers.
This profile is name-resolved from public bibliographic data linked to ORCID 0000-0002-8865-7896; 'Dahua Lin' may correspond to more than one researcher, and 2026 counts may be incomplete.