Peking University researcher building evaluation benchmarks for embodied world models, robot policies, and graph ML
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-08) 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: the OpenAlex record shows an h-index of 0 and only five works indexed in 2026, so this axis — cumulative publication volume, venue quality, and authorship roles — cannot be measured reliably until a fuller record accumulates.
Impact—
Impact is unscored: with h-index and i10-index both at 0 in the OpenAlex snapshot, there is no citation signal yet. This axis would measure citations, citation velocity, and uptake of the proposed benchmarks in follow-up work once data accumulates.
Focus0/100
Focus scores 0 because the record is deliberately spread thin: Evolve's field distribution splits into equal thirds — embodied AI / world-model evaluation, robot learning / embodied AI evaluation, and graph representation learning with domain adaptation — with no dominant field. Topic counts reinforce the breadth: social robot interaction and HRI (3), reinforcement learning in robotics (2), robot manipulation and learning (2), multimodal ML applications (2), topic modeling (1).
Collaboration66/100
Collaboration scores 66, backed by 27 distinct co-authors in the OpenAlex record. All three reviewed papers are middle-author arXiv contributions, and the recurring co-author list spans a dozen names including Xuanyi Liu, Haofeng Wang, Ruiqi Li, and Siwei Ma. The pattern suggests steady teamwork within a Peking University-centered network rather than isolated work, though paper-level roles are not yet mapped.
Activity52/100
Activity scores 52 on a thin but current base: the OpenAlex snapshot records 5 works in the last three years, all indexed in 2026 (countsByYear lists 2026: 5, with no earlier years). All three Evolve-reviewed papers appeared on arXiv in 2026, indicating a concentrated recent burst; the short window means the score should be read as early-career velocity, not sustained long-term productivity.
Reproducibility70/100
Reproducibility scores 70, the mean over three Evolve-reviewed papers. IndustrialVLA-Bench is rated high readiness — a traceable multi-axis protocol with explicit evidence tiers, though based on n=3 runs per setting — while TriWorldBench and the EviGDA graph-domain-adaptation paper are rated medium; Evolve notes TriWorldBench's extracted text contains no results or baselines despite its 500-episode, 19-metric protocol.
Research summary
Jiaqi Zhang is a machine learning researcher at Peking University whose profile, name-resolved from OpenAlex bibliographic data, centers on rigorous evaluation of embodied AI systems. The recent record is compact but coherent: five works in the last three years, all indexed in 2026 in the current snapshot, spread evenly across three strands — embodied world-model evaluation, robot policy benchmarking, and graph representation learning under distribution shift (one third each by Evolve's field distribution). Within Evolve's reviewed catalog, Zhang contributes as a middle author on three arXiv papers: TriWorldBench, a tri-view consistency benchmark for embodied world models spanning 500 episodes and 19 metrics; IndustrialVLA-Bench, a traceable multi-axis evaluation of six vision-language-action and world-action models that finds LIBERO nearly saturated (1.6 pp spread) while robustness and paraphrase axes separate models by 15–31 pp; and EviGDA, which pairs an adapted graph expert with a source-only MLP for graph domain adaptation. Evolve scores reproducibility at 70 across all three reviewed papers, with IndustrialVLA-Bench rated high. Citation indicators (h-index and i10-index of 0) indicate an early-stage record; topic signals — social robot interaction and HRI, reinforcement learning in robotics, and multimodal ML — corroborate the embodied-evaluation focus.
Collaboration pattern
Zhang works in team settings: the OpenAlex record lists 27 distinct co-authors and Evolve assigns a collaboration score of 66. All three reviewed papers are middle-author positions rather than solo or corresponding-author leads, suggesting a contributor role within larger groups. Recurring names include Xuanyi Liu, Haofeng Wang, Ruiqi Li, Danni Yu, Shaofeng Zhang, Zicheng Zhang, and Siwei Ma; the co-author list also contains another 'Jiaqi Zhang', which may reflect same-name entanglement in the OpenAlex graph.
Research keywords
Highlights
TriWorldBench: a tri-view consistency benchmark for embodied world models
IndustrialVLA-Bench separates robot policies where LIBERO is saturated
EviGDA: expert fusion for graph domain adaptation
Timeline
- 2026Released TriWorldBench on arXiv, a 500-episode, 19-metric tri-view consistency benchmark for embodied world models (middle authorship).
- 2026Co-authored EviGDA ('Graph Domain Adaptation Does Not End with Representation Learning'), fusing an adapted graph expert with a source-only MLP.
- 2026Contributed to IndustrialVLA-Bench, a six-model multi-axis evaluation of open VLA/WAM policies with high reproducibility readiness.
- 2026OpenAlex snapshot indexes 5 works for this researcher in this single year, with 27 distinct co-authors across the recent record.
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.
- Bibliometric indicators come from an OpenAlex identity resolution with no linked ORCID; 'Jiaqi Zhang' is a common name and the co-author list itself contains a second 'Jiaqi Zhang', so same-name records may be merged or split.
- Output and impact axes are unscored because h-index and i10-index are both 0 and the snapshot indexes only 2026 works; no citation totals are available.
- All three reviewed papers are 2026 arXiv preprints with middle (non-corresponding) authorship; the input data contains no peer-reviewed venue, funding, or career information.
- The reproducibility score of 70 averages Evolve reviews of only 3 papers; Evolve notes missing results in TriWorldBench's extracted text and n=3 runs in IndustrialVLA-Bench.