Ant Group researcher bridging data-efficient RLVR post-training and multi-agent LLM safety
2
papers here
2
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—
Output cannot be scored yet: no publication count, venue record, or authorship totals are present in the data. This axis would measure total scholarly production — number of papers, publication venues, and career-long output volume.
Impact—
Impact is unscored because citation counts and influence signals are absent from the current snapshot. This axis would track citations, downstream adoption, and how often other work builds on this researcher's contributions.
Focus0/100
Focus scores 0 because the field distribution splits exactly evenly: 0.5 machine learning (LLM post-training / RLVR systems) and 0.5 AI safety and security (LLM agents). With only two reviewed papers, one per area, this reads as an early dual-track profile rather than scattered diffuse work; consolidating one thread or adding papers in either area would sharpen the score.
Collaboration60/100
The collaboration score is 60, computed from a network of 19 distinct coauthors across the reviewed papers. Twelve recurring collaborators — including Shenzhi Yang, Guangcheng Zhu, Kai Tang, Zhengqing Zang, Bo An, and Bo Han — span both tracked areas, consistent with industry team science. Both catalog papers place Xing Zheng in a middle authorship slot without corresponding role, indicating broad but not yet self-led collaboration.
Activity—
Activity lacks underlying data: no yearly publication rhythm or multi-year track record is available from the snapshot. This axis would measure publishing cadence and sustained productivity across consecutive years once a longer bibliographic window exists.
Reproducibility55/100
Reproducibility averages 55 over 2 papers, each assessed at medium repro readiness. DE-Venus is built on the verl framework, which aids reuse, though reviewers note small margins without variance checks; the epidemic-account paper contributes a usable susceptibility benchmark while its contagion dynamics remain untested. Artifacts appear workable but not yet frictionlessly reproducible.
Research summary
Xing Zheng is a researcher at Ant Group (China) whose reviewed work splits evenly across two live frontiers: machine learning for LLM post-training — specifically RLVR systems — and AI safety and security for LLM agents. The Evolve catalog holds two 2026 arXiv papers, both with Xing Zheng in a middle authorship position and neither as corresponding author. The first, DE-Venus, is a verl-based data-efficient RLVR framework touching GRPO-style training under weak supervision and noisy labels; Evolve credits it as well-engineered with real savings while flagging small margins without variance checks and self-cited methods. The second proposes an epidemic account of collective loss of control in multi-agent LLM systems, anchored by a susceptibility benchmark whose contagion dynamics remain untested. Bibliometrically, the coauthor network spans 19 distinct collaborators, twelve of whom recur across both areas. Reproducibility averages 55 across 2 papers, each rated medium readiness. Output, impact, and activity metrics are not yet in the record, so this is an early-stage profile: it depicts an industry researcher contributing inside sizeable teams to making post-training cheaper and multi-agent systems safer, not yet a self-led research line, and the name is resolved from public bibliographic data.
Collaboration pattern
Collaboration is team-based: 19 distinct coauthors appear across the reviewed work, and twelve recur, including Shenzhi Yang, Guangcheng Zhu, Kai Tang, Zhengqing Zang, Haobo Wang, Yingfan Ma, Bowen Song, Bo An, Bo Han, Lei Feng, Weiqiang Wang, and Junbo Zhao. The list spans both tracked areas, and Xing Zheng sits in middle, non-corresponding authorship on both papers — a contributor inside larger industry teams rather than a line leader so far.
Research keywords
Highlights
DE-Venus: data-efficient RLVR built on verl
An epidemic account of multi-agent loss of control
Timeline
- 2026Co-authored DE-Venus, a verl-based data-efficient RLVR framework (arXiv), in a middle, non-corresponding authorship role.
- 2026Co-authored the epidemic account of collective loss of control in LLM agent systems, contributing a susceptibility benchmark (arXiv).
- 2026Published in both tracked fields within the same year — RLVR post-training and agent safety — inside a network of 19 distinct coauthors.
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 derive from public bibliographic data (OpenAlex-based) only; no ORCID link is available, so identity is name-resolved and may be ambiguous for this name.
- Output, impact, and activity dimensions are unscored: citation counts, publication totals, and yearly productivity are absent from the provided data.
- Profile scope is limited to the two Evolve-reviewed 2026 arXiv preprints; peer-review status and work outside the catalog are not reflected.
- Authorship positions in the catalog are middle and non-corresponding, so individual leadership and contribution within teams cannot be assessed from these data.