AI safety researcher at Foshan University probing adversarial robustness in LLMs and wearable activity recognition.
2
papers here
0
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
4
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-27) 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: only two 2026 arXiv preprints are reviewed and yearly OpenAlex counts are sparse, so volume and venue quality cannot be judged. This axis would measure publication volume, venue tier, and consistency.
Impact—
Not yet scored: citation signals are thin (h-index 4, i10-index 2, sparse per-year counts), too limited for a stable influence score. This axis would measure citations, citation velocity, and uptake of the two preprints.
Focus—
Not yet scored: OpenAlex topic labels are spread across memory/neural computing (6), phase-change materials (5), activity recognition (3), and circuit-design topics, so no dominant line emerges. This axis would measure concentration around a core agenda.
Collaboration39/100
Evolve scores collaboration 39/100 from real bibliometric data: 6 distinct coauthors over the last three years, matching a tight pool of Hua Cheng, Zeyu Feng, Yongda Yu, Yuzhe Luo, Yuan Wei, and Renju Liu. Both catalog papers are first-authored but not corresponding, suggesting leadership of experiments rather than supervision of projects. A small, stable team keeps the score low-moderate; wider external collaboration would lift it.
Activity72/100
Activity scores 72/100 on Evolve's radar: 11 works in the last three years per OpenAlex, with yearly counts peaking at 7 in 2024, 3 in 2026 so far, and 1 each in 2023 and 2025. Two first-authored arXiv preprints in 2026 sustain a visible cadence, though the lone 2005 record and thin 2025 output make the trajectory uneven rather than steadily cumulative.
Reproducibility—
Not yet scored: Evolve reviewed only two papers, both marked medium reproducibility readiness — a base too small for an axis score. This axis would measure code, data, and artifact release across the full output.
Research summary
Qingyu Wu is a researcher at Foshan University, China, whose Evolve-reviewed work centers on adversarial robustness: auditing machine-learning systems under weak or adversarial signal conditions. Two first-authored 2026 arXiv preprints anchor this profile. ENDOPROMPT, in LLM security, learns utility-degrading prefixes from unlabeled victim signals; Evolve reports a -26.8pp utility effect across 27/28 evaluation cells and calls the work honest, while noting its request-matching advantage remains unproven. The second paper reframes temporal perturbations in wearable activity recognition as sensor-like biases, introducing a label-sealed select-then-project audit that finds failures mostly DC-carried — rigorous, though Evolve flags circularity risk in DC-retained search. The OpenAlex snapshot lists h-index 4, i10-index 2, 11 works in the last three years, and 6 distinct coauthors, with Evolve scoring activity 72/100 and collaboration 39/100. Counted output clusters in 2024 (7 works) and 2026 (3 so far), after a lone 2005 record and a quiet 2025. Topic labels span advanced memory and neural computing (6), phase-change materials and chalcogenides (5), and context-aware activity recognition (3), broader than the reviewed AI-safety papers — possibly multi-area work or name-resolution noise. This remains an early-stage, preprint-heavy profile.
Collaboration pattern
Collaboration is compact: 6 distinct coauthors in the last three years — Hua Cheng, Zeyu Feng, Yongda Yu, Yuzhe Luo, Yuan Wei, and Renju Liu — matching Evolve's 39/100 collaboration score. Both catalog papers are first-authored without corresponding-author status, consistent with leading experiments inside a small stable group rather than supervising projects. No cross-institution patterns are visible in the provided data.
Research keywords
Highlights
ENDOPROMPT: learning utility-degrading prefixes from unlabeled victim signals
Label-free auditing shows wearable HAR perturbation failures are mostly DC-carried
Timeline
- 2005OpenAlex indexes a single work in 2005; no further indexed output appears until 2023.
- 2023Indexed output resumes with 2 works recorded in 2023.
- 2024Peak indexed output: 7 works recorded in 2024.
- 2025Indexed output drops to 1 work in 2025.
- 20263 works counted so far, including two first-authored arXiv preprints reviewed by Evolve on LLM security and wearable-sensing robustness.
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 come from an OpenAlex snapshot only; identity is name-resolved from public bibliographic data, and Qingyu Wu is a common name, so the record may aggregate namesakes.
- OpenAlex topic labels (memory/neural computing, phase-change materials, VLSI) differ from the adversarial-robustness focus of the reviewed papers, reflecting either multi-area work or name-resolution noise.
- Only two Evolve-reviewed papers exist, both 2026 arXiv preprints; output, impact, focus, and reproducibility axes therefore carry no scores.
- The single 2005 work followed by a gap to 2023 makes career-stage and trend inference unreliable.