Jian Liu — AI for science and speech at Shanghai AI Lab: MoE speaker verification and LLM-based omics benchmarking
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) 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 yet: this snapshot lacks publication-count and venue data for Jian Liu. The axis would measure total scholarly output — the number of publications, the venues they appear in, and their distribution over years — none of which is currently available in the underlying bibliometric data.
Impact—
No impact score yet: citation counts and downstream uptake are absent from the underlying data. The axis would measure how much later work builds on these publications, which cannot be estimated or scored from the two reviewed 2026 preprints alone.
Focus0/100
The focus axis scores 0, the lowest concentration, because the field distribution is an even 50/50 split between computational biology (AI for science) and speech processing / speaker recognition, with no dominant field across the two reviewed papers. This is a dual-domain pattern rather than a deficiency: both lines involve deep architecture engineering — MoE CNN-Transformer models and distilled tool-augmented LLMs — but the profile cannot yet identify a single primary field.
Collaboration59/100
Collaboration scores 59, a mid-range value based on 18 distinct coauthors across the two catalogued papers. Both papers list Liu in a middle author position without corresponding-author status, consistent with work inside larger teams. Twelve collaborators recur in the network, including Jie Ying, Zhefan Wang, Wanli Ouyang, Stan Z. Li and Fan Yang, indicating stable internal partnerships alongside a moderately sized but not yet broad external network.
Activity—
No activity score yet: yearly publication volume and cadence are not available in this snapshot beyond the two 2026 papers. The axis would measure research tempo — outputs per year and their consistency over time — and requires a longer publication window than the current record provides.
Reproducibility78/100
Reproducibility scores 78, the mean over the two assessed papers (reproPapers: 2). The omics benchmark paper rates high readiness with strong releases, while the MECT speaker-verification paper rates medium: its VoxCeleb minDCF and 100 ms streaming results are single-run and untested for statistical significance. A 78 across two 2026 preprints suggests a generally reusable artifact base, with significance testing as the main gap.
Research summary
Jian Liu is a researcher at Shanghai Artificial Intelligence Laboratory whose catalogued work spans two distinct areas: speech processing / speaker recognition, and computational biology in the AI-for-science mode. The Evolve catalog holds two 2026 arXiv preprints, in both of which Liu is a middle-listed, non-corresponding author — a contributor role within larger teams. The first, MECT, proposes a mixture-of-experts CNN-Transformer network for speaker verification, claimed as the first fully supervised MoE-based SV model with state-of-the-art VoxCeleb minDCF and 100 ms streaming inference; Evolve judged the results strong but single-run and untested for significance, at medium reproducibility readiness. The second introduces tool-augmented on-policy distillation (TA-OPD) for LLM domain adaptation in sequence-based omics tasks, together with what Evolve calls the first omics reasoning benchmark; it rates high on reproducibility readiness, though LLM-judge circularity and small-sample statistics temper its claims. Bibliometrically, the network contains 18 distinct coauthors with 12 recurring collaborators, and the field distribution is an even 50/50 split between the two areas — hence a focus score of 0, a collaboration score of 59, and a reproducibility mean of 78 across the two scored papers. Output, impact and activity axes remain unscored for lack of data.
Collaboration pattern
Jian Liu's catalogued work is team-based: both 2026 arXiv papers carry middle authorship without corresponding status, and the coauthor network holds 18 distinct collaborators. Twelve recur frequently — among them Jie Ying, Zhefan Wang, Wanli Ouyang, Stan Z. Li and Fan Yang — indicating a stable core group alongside contributors spanning both the speech and the AI-for-science lines. The collaboration score of 59 reflects a moderately sized, internally consistent network rather than a broad external one.
Research keywords
Highlights
First omics reasoning benchmark with tool-augmented on-policy distillation
MoE CNN-Transformer pushes streaming speaker verification
Timeline
- 2026Co-authored MECT, a mixture-of-experts CNN-Transformer network for speaker verification, released as an arXiv preprint with middle authorship and no corresponding role.
- 2026Co-authored the tool-augmented on-policy distillation (TA-OPD) paper for LLM domain adaptation in sequence-based omics tasks, introducing the first omics reasoning benchmark, again as a middle-listed arXiv author.
- 2026Released reviewed work in both tracked fields within the same year — speech processing / speaker recognition and computational biology (AI for science), each accounting for half of the catalogued output.
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.
- Identity is name-resolved from public bibliographic data only: 'Jian Liu' is a common name, and no ORCID, country, or disambiguating identifier is provided in this snapshot.
- Output, impact and activity dimensions are unscored because citation counts, venue metrics and yearly publication volume are absent from the underlying bibliometric data.
- The Evolve catalog covers only two 2026 arXiv preprints with middle, non-corresponding authorship; peer-reviewed or lead-author work is not represented in this profile.
- Radar scores rest on a small base — two papers, 18 distinct coauthors and a 50/50 field split — so the collaboration, focus and reproducibility values may shift as the catalog grows.