UNSW Sydney researcher in neural architecture search and diffusion watermarking robustness.
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
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 catalog currently holds only three 2026 arXiv papers, and no aggregated output metrics such as total publication count or venue record are available. This axis would measure overall publication volume and productivity over time.
Impact—
Not yet scored: citation counts, h-index, and downstream adoption are unavailable in the current data. This axis would measure how often the three reviewed papers are cited and built upon by follow-up work.
Focus8/100
The focus score is 8/100, and the field distribution of the three cataloged papers splits 66.7% neural architecture search and 33.3% AI security on diffusion watermark removal. The cataloged output therefore spans two distinct directions — efficiency-oriented NAS performance prediction and adversarial analysis of generative model watermarking — rather than one consolidated thread, so topic concentration is low at this stage.
Collaboration41/100
The collaboration score is 41/100, with 7 distinct frequent coauthors in the snapshot, including Xiaoyu Li, Jiaojiao Jiang, and Yang Song. All three reviewed papers place the researcher in a middle author position without corresponding status, suggesting contribution within team-driven projects rather than sole or leading authorship; the co-author network remains modest in size so far.
Activity—
Not yet scored: year-by-year publishing activity cannot be assessed because the catalog contains only 2026-dated papers and no longitudinal metadata. This axis would measure publishing cadence and sustained research output across recent years.
Reproducibility55/100
Reproducibility readiness averages 55 across the three reviewed papers (reproPapers = 3), each rated medium. Evolve's review of RiPPLE notes that no code release is stated, while CoRA-NAS and DRIFT are judged largely supported by reported evidence — though DRIFT relies on a single backbone and lacks significance tests — leaving room for stronger artifact availability.
Research summary
Zheng Gao is a researcher at UNSW Sydney whose reviewed work to date sits at the intersection of efficient deep learning and AI security. This full profile is grounded in three 2026 arXiv papers reviewed by Evolve, in all of which the researcher holds a middle author position without corresponding status. Two papers advance neural architecture search: RiPPLE predicts whole-space architecture rankings from early training signals by combining zero-cost proxies with learning-curve extrapolation, while CoRA-NAS proposes a label-free two-stage scheme of coarse ranking plus anchor-residual refinement, with reported robustness across four vision benchmarks. The third, DRIFT, is a black-box attack that removes diffusion model watermarks by deflecting the generative trajectory, reporting 98-100% success across nine watermarking schemes. Bibliometric signals in the snapshot include 7 distinct frequent coauthors, a collaboration score of 41, a field split of roughly two-thirds NAS to one-third AI security, and a reproducibility readiness mean of 55 across the three papers, all rated medium. Publication volume, citation impact, and longitudinal activity are not yet scored.
Collaboration pattern
The snapshot lists 7 distinct frequent coauthors — Xiaoyu Li, Jiaojiao Jiang, Yifan Yang, Zhaoyan Wang, Rui Bao, Xiaoyan Feng, and Yang Song. Across all three reviewed 2026 papers the researcher appears in middle author positions without corresponding status, a pattern consistent with specialist contributions inside small collaborative teams spanning both the NAS and AI-security lines.
Research keywords
Highlights
RiPPLE: cross-space NAS ranking predicted from early training
CoRA-NAS: label-free two-stage ranking with anchor-residual refinement
DRIFT: black-box diffusion watermark removal via trajectory deflection
Timeline
- 2026Publishes RiPPLE, proposing cross-space NAS performance prediction from early training signals.
- 2026Publishes CoRA-NAS, a label-free two-stage coarse ranking and anchor-residual refinement method for NAS.
- 2026Publishes DRIFT, a black-box attack removing diffusion model watermarks by deflecting the generative trajectory.
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 dimensions are computed from a limited OpenAlex-based snapshot; citation counts, h-index, and cumulative output totals are not available.
- The profile is name-resolved from public bibliographic data; 'Zheng Gao' is a common name and no ORCID is provided, so identity matching carries residual uncertainty.
- All paper-level evidence comes from three 2026 arXiv preprints reviewed by Evolve; peer-reviewed venue acceptance is unconfirmed.
- Reproducibility scores and verdicts reflect Evolve's desk reviews of the papers, not independent code execution or replication.