Yu Zheng

Ant Group (China)

Unclaimed · AI-structured profileOpenAlex

Yu Zheng — Ant Group researcher spanning LLM agent security, speaker verification, and text anomaly detection

papers here

reviewed by Evolve

first-author

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 scored — insufficient data. Only three 2026 arXiv preprints are on record for this name, with no full publication count or venue history. This axis would measure total output volume, venue tier, and author position over a career.

Impact—

Not scored — insufficient data. No citation counts, h-index, or venue-level influence metrics are present in the input. This axis would measure citations and uptake of the speaker-verification, anomaly-detection, and LLM-agent-security work over time.

Focus0/100

Scores 0/100: the field distribution is an exact three-way split, with computer security (LLM agent security / applied cryptography), speech processing / speaker recognition, and machine learning (text anomaly detection) each covering one third of the record. No direction dominates. With only three 2026 arXiv papers on file, this may reflect genuine cross-domain range or an agenda still taking shape rather than scattered effort.

Collaboration50/100

Scores 50/100, derived from 11 distinct coauthors across the three 2026 papers. The pattern is team-based rather than solo: Jian Liu is listed more than once among frequent coauthors, alongside Qizhi Zhang, Jinghan Peng, ChangHao Zhang, Weiqiang Wang, Kehan Yan, Yue Tan, Qingfeng Chen, Shiyuan Li, and Yixin Liu. Each paper assembles its own collaborator group, consistent with product-team research inside Ant Group across three separate fields.

Activity—

Not scored — insufficient data. All three papers date to 2026 and no longer-term publication history is available, so yearly cadence cannot be judged. This axis would measure publication rhythm and research continuity across multiple years.

Reproducibility55/100

Scores 55/100, the mean reproducibility-readiness rating across all three reviewed papers (3 of 3 rated, all medium). The rating tracks Evolve's specific reservations: MECT's state-of-the-art VoxCeleb minDCF and 100 ms streaming claims come from single runs with no significance testing; SIM reports strong gains on three datasets but likewise omits significance tests and sits near-floor on AUPRC; KITA is judged a sound design whose motivating prompt-injection threat is never actually attacked.

Research summary

Yu Zheng is a researcher at Ant Group whose 2026 record spans three distinct areas: speech processing and speaker recognition, machine learning for text anomaly detection, and computer security focused on LLM agent security and applied cryptography. Evolve's catalog holds three arXiv preprints, two of them first-authored. MECT proposes a Mixture-of-Experts CNN-Transformer network for speaker verification, with claimed state-of-the-art VoxCeleb minDCF and 100 ms streaming inference, though results are single-run and untested for significance. SIM, a co-authored work, introduces subspace interaction with hard pseudo-anomalies for token-level text anomaly detection, reporting strong gains on three datasets without significance tests. KITA, first-authored, enforces review before LLM agent authorization through t-of-n threshold BLS signatures with keys isolated from all LLMs; Evolve judges the design sound but notes the motivating prompt-injection threat is never attacked. Radar scores put reproducibility at 55/100 across three rated papers and collaboration at 50/100 with 11 distinct coauthors, while focus scores 0/100 given an even three-way field split. Output, impact, and activity axes lack data. The profile is name-resolved from public bibliographic data; 'Yu Zheng' is a common name.

Collaboration pattern

Collaboration is team-based across three distinct research clusters. Eleven distinct coauthors appear in the record, and Jian Liu is listed more than once among frequent coauthors, suggesting repeat collaboration; other recurring names include Qizhi Zhang, Jinghan Peng, ChangHao Zhang, Weiqiang Wang, Kehan Yan, Yue Tan, Qingfeng Chen, Shiyuan Li, and Yixin Liu. All three papers are 2026 arXiv preprints, consistent with group research inside Ant Group rather than solo output.

Research keywords

speaker verificationmixture of expertsCNN-Transformer architecturestreaming inferencetoken-level text anomaly detectionsubspace analysispseudo-anomaly generationLLM agentsprompt injectionthreshold signaturesBLS signatureskey isolation

Highlights

MECT: claimed-first MoE-based fully supervised speaker verification with SOTA VoxCeleb results
First-author 2026 arXiv paper proposing a Mixture-of-Experts CNN-Transformer network for speaker verification. Evolve's review notes claimed state-of-the-art VoxCeleb minDCF and 100 ms streaming inference, while flagging single-run results and the absence of significance testing. Reproducibility readiness: medium.
SIM: subspace interaction for token-level text anomaly detection
Co-authored 2026 arXiv paper combining subspace attention with hard pseudo-anomaly generation for token-level anomaly detection in text. Evolve reports strong gains on three datasets, but notes the absence of significance tests and near-floor AUPRC. Reproducibility readiness: medium.
KITA: key-isolated threshold signing to gate LLM agent authorization
First-author 2026 arXiv paper enforcing review-before-authorization for LLM agents via t-of-n threshold BLS signatures, with key material isolated from all LLMs. Evolve judges the design sound but notes the motivating prompt-injection threat is never actually attacked. Reproducibility readiness: medium.

Timeline

2026First-authored MECT, a Mixture-of-Experts CNN-Transformer model for speaker verification claiming SOTA VoxCeleb minDCF and 100 ms streaming inference.
2026Co-authored SIM, a subspace-interaction method for token-level text anomaly detection with strong reported gains on three datasets.
2026First-authored KITA, key-isolated t-of-n threshold BLS signing for enforcing review before LLM agent authorization.

Papers in the Evolve catalog

2026-09-21first author
MECT: Mixture of Experts with CNN-Transformer Network for Speaker verification
2026-09-08
SIM: Subspace Interaction-based Method for Token-Level Text Anomaly Detection
2026-09-05first author
From Review to Authorization: Key-Isolated Threshold Signing for LLM Agents

Frequent co-authors

Qizhi Zhang×1 Jinghan Peng×1 ChangHao Zhang×1 Jian Liu×1 Weiqiang Wang×1 Jian Liu×1 Kehan Yan×1 Yue Tan×1 Qingfeng Chen×1 Shiyuan Li×1 Yixin Liu×1

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 radar scores are computed from OpenAlex-derived data only; citation counts, h-index, and venue-level metrics are absent from this snapshot.
'Yu Zheng' is a common Chinese name; this profile is name-resolved from public bibliographic data and the three arXiv papers reviewed by Evolve.
Only three 2026 arXiv preprints are on record; no peer-reviewed venue, long-term publication history, or career details were available.
Output, impact, and activity dimensions are left unscored because their underlying metrics are missing from the input, not because performance is weak.