Zhengyao Jiang — LLM agents researcher: autonomous research agents and agentic pre-training data curation
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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 scorable: this profile's evidence base is only two reviewed 2026 arXiv papers, with no cumulative publication counts in the snapshot. This axis would measure total papers, venues, and yearly output volume, none of which the current data provides.
Impact—
No citation or uptake data is present in the input, so impact is unscored. This axis would measure citations, referencing works, and downstream adoption of the recursive self-improvement and data-selection methods described in the catalog.
Focus0/100
Focus scores 0 because the cataloged output is split exactly evenly: the field distribution assigns 0.5 to autonomous research agents / LLM agent systems and 0.5 to LLM pre-training data curation. With two papers, one per field, there is no dominant theme yet; both lines share an agentic-search methodology, so concentration may emerge as more work is cataloged.
Collaboration41/100
The collaboration dimension scores 41 from 7 distinct coauthors across the two cataloged 2026 papers. Recurring collaborators include Dhruv Srikanth, Bingchen Zhao, Yuxiang Wu, Dixing Xu, and Yan Meng. Jiang holds senior ordering positions — last author on the recursive self-improvement study (non-corresponding) and a middle author on AutoData — a pattern consistent with small, specialist LLM-lab teams rather than large multi-institution consortia.
Activity—
Recent-activity signals cannot be computed from two papers dated 2026 alone. This axis would measure publication cadence and research continuity across recent years, which the current snapshot does not cover; no per-year output series is available yet.
Reproducibility38/100
Reproducibility scores 38, the mean readiness rating across the two cataloged papers. The last-authored recursive self-improvement study is rated low: Evolve notes it rests on few runs and noisy grade comparisons despite the 0.703→0.778 gain. AutoData is rated medium: val-bpb improvements hold across scales, but CORE gains lack statistical support. Jiang is not the corresponding author on either paper.
Research summary
Zhengyao Jiang is an early-stage researcher at the intersection of LLM agent systems and pre-training data curation, with two 2026 arXiv papers in the Evolve catalog. The first, Recursive self-improvement of AI research agents, is a last-authored (non-corresponding) study of AIDE², an eight-day self-improvement loop that lifts an AI research agent's grade from 0.703 to 0.778 with held-out transfer — compelling, in Evolve's view, but resting on few runs and noisy grade comparisons. The second, AutoData: Agentic Search for Pre-training Data Selection, is a middle-authored collaboration in which LLM-agent search over executable data-selection recipes beats curated-pipeline baselines on val-bpb across scales, though CORE gains lack statistical support. The catalog splits evenly (0.5/0.5) between autonomous research agents and LLM pre-training data curation, so the focus dimension scores 0. Radar scores collaboration 41 (7 distinct coauthors) and reproducibility 38 (mean readiness across two papers rated low and medium); output, impact, and activity lack sufficient data and are unscored. No institution, country, or ORCID accompanies the record, so the profile is name-resolved from public bibliographic data.
Collaboration pattern
Across the two cataloged 2026 papers Zhengyao Jiang works with 7 distinct coauthors, including recurring collaborators Dhruv Srikanth, Bingchen Zhao, Yuxiang Wu, Dixing Xu, and Yan Meng. Author-order roles differ — last author (non-corresponding) on the recursive self-improvement study, middle author on AutoData — suggesting involvement in both direction-setting and core technical work within small, focused LLM research teams. The radar collaboration score is 41.
Research keywords
Highlights
Eight-day recursive self-improvement lifts AI research agent grade from 0.703 to 0.778
Agentic search over executable data-selection recipes beats curated pre-training pipelines
Timeline
- 2026Last-authored (non-corresponding) arXiv paper on recursive self-improvement of AI research agents, reporting an agent-grade gain from 0.703 to 0.778 over an eight-day loop.
- 2026Middle-authored arXiv paper AutoData applies agentic search to LLM pre-training data selection, outperforming curated-pipeline baselines on val-bpb across scales.
- 2026Cataloged output for the year splits evenly between autonomous research agents and pre-training data curation, with reproducibility readiness rated low and medium respectively.
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.
- Radar scores are computed by Evolve from a thin bibliometric snapshot (two reviewed papers); output, impact, and activity axes lack sufficient data and are unscored.
- No institution, country, or ORCID is attached to this record; the profile is name-resolved from public bibliographic data, so identity ambiguity is possible.
- Both papers are 2026 arXiv preprints; peer-reviewed venue status is not captured in the provided data.
- Collaborator names come from the small cataloged sample (7 distinct coauthors) and may not reflect the full collaboration network.