AI researcher at Beihang University bridging agentic visual generation and LLM post-training, with rising output.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
71
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 yet scored: the snapshot reports an h-index (71) and i10-index (173) but lacks the complete work counts and citation aggregates this axis needs, so total publication volume cannot be computed yet.
Impact—
Not yet scored: no citation totals, field-normalized percentiles, or per-paper citation distributions are available in the snapshot, so this axis — which would measure scholarly influence relative to field peers — cannot be estimated.
Focus0/100
Scores 0/100, flagging very low concentration: fieldDist splits evenly (0.5/0.5) between agentic visual generation and LLM agents/post-training, and OpenAlex topic clusters span five areas — Advanced Neural Network Applications (122), Domain Adaptation and Few-Shot Learning (82), Software Testing and Debugging (46), Security and Verification (31), and Visual Attention and Saliency Detection (31). The record is deliberately broad rather than concentrated in one niche.
Collaboration75/100
Scores 75/100. The recent three-year window records 71 works with 43 distinct coauthors, and 12 frequent collaborators recur across projects — including named individuals (Xu Xu, Jinxiu Liu, Zhangbo Qiao) and the collective 'NeoHorse Team'. On both reviewed 2026 papers Zhang is a middle author and not corresponding, consistent with a contributor role inside larger, team-based research groups rather than solo-led projects.
Activity100/100
Scores 100/100, the strongest axis. Yearly counts climb from single digits through 2014 (2) to 8 in 2015, 16 in 2018, 30 in 2020, and a peak of 45 in 2023, followed by 39 (2024), 27 (2025), and 5 recorded so far in 2026. The most recent three-year window alone holds 71 works, indicating sustained, currently high publishing throughput.
Reproducibility38/100
Scores 38/100, computed from Evolve's review of 2 papers (reproMean 38). OmniHarness is rated medium: its symbolic-policy claim of 95.0% Creative resolution on ComfyBench is judged evidence-sufficient, but numerical consistency needs verification. NeoHorse-1 is rated low: evaluations are good, yet the key curriculum claim lacks a random-order ablation. Reproducibility evidence is thin at two papers.
Research summary
Xiangyu Zhang (Beihang University, China; ORCID 0000-0003-2138-4608) presents an OpenAlex-resolved profile split evenly between agentic visual generation and NLP for LLM agents and post-training (0.5 each). The snapshot reports an h-index of 71 and an i10-index of 173. Output has accelerated sharply: yearly counts rise from 8 (2015) to 16 (2018), 30 (2020), and a peak of 45 (2023), with 39 in 2024 and 27 in 2025, and 71 works in the most recent three-year window across 43 distinct coauthors. Evolve scores this record 100/100 for activity and 75 for collaboration, while focus scores 0, reflecting five dispersed topic clusters: Advanced Neural Network Applications (122), Domain Adaptation and Few-Shot Learning (82), Software Testing and Debugging (46), Security and Verification (31), and Visual Attention and Saliency Detection (31). Evolve has reviewed two 2026 arXiv papers with Zhang as a middle author: OmniHarness claims symbolic policy learning lifts visual-generation agents to 95.0% Creative resolution on ComfyBench (evidence sufficient; numerical consistency to verify; medium reproducibility readiness), and NeoHorse-1 claims routing-harness data with curriculum learning and on-policy distillation lifts 4B/9B macro averages to 64.87/69.04 (good evaluations; the curriculum claim lacks a random-order ablation; low readiness). Output and impact axes remain unscored pending fuller citation aggregates.
Collaboration pattern
The profile shows team-based collaboration: 43 distinct coauthors appear across the 71 works of the recent three-year window, with 12 frequent collaborators including Xu Xu, Jinxiu Liu, Zhangbo Qiao, Jiaxing Lu, and the collective author 'NeoHorse Team'. On both Evolve-reviewed 2026 papers Zhang is a middle author and not corresponding, suggesting contribution within larger group efforts spanning agentic visual generation and LLM post-training projects.
Research keywords
Highlights
Symbolic policy learning for generalizable visual generation agents
Recursive self-improvement via agentic post-training
Sharply accelerating publication trajectory
Recorded yearly counts rise from single digits before 2015 to a 2023 peak of 45 (39 in 2024, 27 in 2025), with 71 works in the most recent three-year window across 43 distinct coauthors; Evolve scores activity 100/100.
Timeline
- 2003Earliest records in the OpenAlex snapshot appear in 2003, with 4 works recorded that year.
- 2015Output leaves the single-digit years behind: 8 works recorded in 2015, up from 2 in 2014.
- 20202019–2020 show 23 then 30 works, roughly triple earlier levels and a steep climb toward the peak.
- 2023The yearly series peaks at 45 works in 2023, followed by 39 in 2024 and 27 in 2025.
- 2026Two Evolve-reviewed arXiv papers appear — OmniHarness and NeoHorse-1 — with 5 works recorded so far in 2026.
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 resolved from OpenAlex only; 'Xiangyu Zhang' is a common name, so this profile is name-resolved from public bibliographic data — verify identity against ORCID 0000-0003-2138-4608.
- Output and impact axes are unscored: the snapshot lacks citation totals and complete work-count aggregates.
- Reproducibility is based on only 2 Evolve-reviewed papers (reproMean 38); the rest of the record has not been reviewed.
- 2026 counts are partial (5 works recorded to date), so the most recent year is incomplete.