Early-career ML researcher at Tsinghua spanning embodied-AI benchmarking, LLM distillation, and graph domain adaptation
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—
No output score is available: this OpenAlex record lacks usable productivity aggregates, so the axis is left unscored. It would measure total works, first/corresponding-author share, and venue mix over time.
Impact—
No impact score is available: the record reports h-index 0 and i10-index 0 with no citation aggregation, typical for a fresh preprint-heavy profile. This axis would measure citations, h-index trajectory, and uptake of the 2026 preprints.
Focus0/100
The focus score is 0 because recorded output splits evenly across three fields — robot learning/embodied-AI evaluation, LLM post-training/distillation, and graph representation learning/domain adaptation — each at exactly one third of the field distribution. Topic tags (Topic Modeling ×3, Advanced Neural Network Applications ×2, Natural Language Processing Techniques ×2) reinforce the breadth-over-depth pattern, which reads as deliberate early-stage exploration rather than a settled single-field agenda.
Collaboration60/100
The collaboration score is 60/100, built on 19 distinct coauthors in the record. The frequent-coauthor list (Yiqi Wang, Zhifeng Rao, Xiaoyang Li, Zhangkai Wu, Yiqun Duan, Mingkai Zheng, Fei Wang, Shan You, Taotao Cai, Zhinan Hou) shows recurring small-team patterns across all three research threads, yet no single long-running partnership dominates, and all three catalog papers are middle-authored rather than led.
Activity52/100
Activity scores 52/100. All recorded output is concentrated in one year: OpenAlex lists 5 works in 2026, and the trailing three-year window contains the same 5 works (recent3yWorks = 5), so there is no multi-year baseline to smooth against. This burst pattern fits a record that was recently linked or a preprint-first output style, and it should be re-read once a second publication year appears.
Reproducibility70/100
Reproducibility is the strongest scored axis at 70/100, averaged across the 3 papers Evolve has reviewed. IndustrialVLA-Bench is rated high readiness with traceable, evidence-tiered protocols; the on-policy distillation study and EviGDA both sit at medium, reflecting single-run or ensembling-heavy result structures. The mean is already solid for a preprint-only record, though it rests on a three-paper sample.
Research summary
Jiaqi Zhang (Tsinghua University) is an early-career machine learning researcher whose visible work splits evenly across three threads: robot learning and embodied-AI evaluation, LLM post-training and distillation, and graph representation learning with domain adaptation — each exactly one third of the OpenAlex field distribution. The Evolve catalog holds three 2026 arXiv preprints, all with middle authorship. IndustrialVLA-Bench delivers a traceable multi-axis evaluation of open vision-language-action and world-action models, finding clean LIBERO performance nearly saturated (1.6 pp spread) while robustness and paraphrase axes separate models by 15–31 pp. An empirical on-policy distillation study shows eight selected hard examples can match full-dataset distillation, with data selection as the decisive factor. EviGDA claims graph domain adaptation SOTA by fusing an adapted graph expert with a source-only MLP, which Evolve reads as principled ensembling rather than new adaptation machinery. OpenAlex records 5 works in 2026, 5 works over the last three years, and 19 distinct coauthors; topic tags include Topic Modeling (3) and Advanced Neural Network Applications (2). Reproducibility averages 70/100 across the three reviewed papers — the strongest scored axis — while output and impact remain unscored because citation data (h-index and i10-index both 0) is not yet available.
Collaboration pattern
With 19 distinct coauthors on record and a collaboration score of 60/100, Jiaqi Zhang works in small rotating teams rather than one fixed group. Recurring names include Yiqi Wang, Zhifeng Rao, Xiaoyang Li, Zhangkai Wu, Yiqun Duan, Mingkai Zheng, Fei Wang, Shan You, Taotao Cai, and Zhinan Hou, and the roster spans all three research threads. All three catalog papers carry middle authorship with no corresponding-author role, consistent with a contributing position.
Research keywords
Highlights
Traceable multi-axis benchmark separates open robot policies where LIBERO saturates
Eight selected hard examples rival full-dataset on-policy distillation
Graph domain adaptation pushed beyond representation learning
Timeline
- 2026Co-authored the IndustrialVLA-Bench preprint on arXiv: a traceable multi-axis evaluation of open robot policy models (middle authorship).
- 2026Co-authored the empirical on-policy distillation study on data efficiency and data selection (arXiv, middle authorship).
- 2026Co-authored EviGDA on graph domain adaptation beyond representation learning (arXiv, middle authorship).
- 2026OpenAlex records 5 works in 2026 — the entire visible output to date — spread evenly across three research fields.
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 derived from OpenAlex only; no ORCID is linked, and 'Jiaqi Zhang' is a common name, so this profile is name-resolved from public bibliographic data and may merge or split identities.
- Output and impact dimensions are missing: the record reports h-index 0 and i10-index 0, reflecting absent citation coverage rather than measured zero impact.
- All three catalog papers are 2026 arXiv preprints; peer-review status, final versions, and publication venues are not reflected.
- The reproducibility score averages Evolve reviews of only 3 papers, so it is sensitive to small sample size.