Alibaba researcher advancing open-ended LLM agent RL via credit propagation and rubric-grounded self-evolution
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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: this snapshot carries no publication-count or venue-volume data for output. The axis would measure how many papers the researcher produces and how consistently across venues and years.
Impact—
Not scored: no citation counts, h-index, or influence metrics are available in this snapshot. Impact would measure how widely the reviewed agent-RL methods are cited, adopted, or built upon by other teams.
Focus100/100
fieldDist shows all catalogued work in a single field, machine learning (LLM agents and reinforcement learning), yielding a maximal focus score of 100. Both 2026 papers — ArenaFlow and ARISE-RL — target open-ended agent RL, one on hierarchical credit assignment via tournament ranking, the other on rubric-grounded self-evolution, so the portfolio is tightly concentrated rather than spread across topics.
Collaboration62/100
A collaboration score of 62 is computed from 22 distinct coauthors in the snapshot. Both reviewed papers place Zheng-Jun Zha in the last author slot of large team rosters, with repeat collaborators including Ruixue Ding, Boli Chen, Pengjun Xie, Qiang Zhang, Kaipeng Zhang, and Jiawei Liu — a stable industrial lab pattern in which the senior author oversees direction rather than hands-on execution.
Activity—
Not scored: the snapshot lacks time-series publication data, so recent-year output cadence cannot be computed. Activity would measure publication frequency and momentum over recent years, beyond the two 2026 preprints reviewed here.
Reproducibility55/100
Reproducibility averages 55 across one assessed paper (reproPapers: 1), and both reviewed preprints are rated medium on reproducibility readiness. ArenaFlow reports a +19.5 gain over ArenaRL on Open-Travel but is flagged for judge circularity and absent variance tests; ARISE-RL's strong gains rest on self-built benchmarks. Releasing code, data, and seed-level variance reporting would lift this axis.
Research summary
Zheng-Jun Zha is an Alibaba Group researcher working at the intersection of large language model agents and reinforcement learning. In the Evolve-reviewed catalog he appears as last author on two 2026 arXiv preprints, a position consistent with senior, supervisory involvement rather than first-author execution; neither paper lists him as corresponding author. The first, ArenaFlow, extends tournament-based trajectory ranking for open-ended agent RL into a hierarchical credit-propagation scheme that reaches down to step- and skill-level rewards, lifting performance by +19.5 over ArenaRL on the Open-Travel benchmark, though Evolve's review flags judge circularity and missing variance testing. The second, ARISE-RL, reports state-of-the-art open-ended agent RL through rubric-grounded generator-solver co-evolution with reward-gated self-distillation; gains are strong but measured on self-built benchmarks. The radar shows maximal focus (100): every catalogued paper falls in a single field, machine learning (LLM agents and reinforcement learning). Collaboration scores 62 on 22 distinct coauthors, with recurring partners including Ruixue Ding, Boli Chen, Pengjun Xie, Kaipeng Zhang, and Jiawei Liu. Reproducibility averages 55 over one assessed paper, and both preprints carry medium reproducibility readiness. Output, impact, and activity axes are unscored because the underlying bibliometric data is not yet available; this profile is therefore a deliberately narrow, evidence-bounded view of a concentrated industrial research line.
Collaboration pattern
The snapshot records 22 distinct coauthors, and the frequent-collaborator list — Ruixue Ding, Boli Chen, Shihang Wang, Pengjun Xie, Xi Chen, Qiang Zhang, Fanrui Zhang, Yinfeng Huang, Yi Zheng, Kaipeng Zhang, Jiawei Liu — reads like a standing industrial lab roster. On both reviewed papers Zheng-Jun Zha sits in the last author position without corresponding-author designation, consistent with senior supervision of a large, stable Alibaba team (collaboration score: 62).
Research keywords
Highlights
ArenaFlow: hierarchical credit propagation for open-ended agent RL
ARISE-RL: rubric-grounded self-evolving agents
A single-track, senior-author research line in agent RL
Both catalogued papers share one field — machine learning (LLM agents and reinforcement learning) — and place Zheng-Jun Zha in the last author slot, giving a focus score of 100 and a concentrated, supervision-oriented profile rather than a spread of topics.
Timeline
- 2026The Evolve catalog records ArenaFlow, a 2026 arXiv preprint with Zheng-Jun Zha as last author, extending tournament-based agent RL with hierarchical credit propagation (+19.5 over ArenaRL on Open-Travel).
- 2026ARISE-RL enters the same 2026 catalog: rubric-grounded generator-solver co-evolution with reward-gated self-distillation, claiming state-of-the-art open-ended agent RL on self-built benchmarks.
- 2026Both 2026 preprints fall in the single field of machine learning (LLM agents and reinforcement learning), with Zheng-Jun Zha in the last author slot on each — the basis for the maximal focus score of 100.
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 scores come from OpenAlex-derived data only; citation counts and publication-volume metrics were not part of this snapshot.
- This profile is backed only by two 2026 arXiv preprints in the Evolve review catalog; no earlier publications or peer-reviewed venue records are included.
- The reproducibility score (55) averages over a single assessed paper, so that axis is statistically thin.
- This profile is name-resolved from public bibliographic data (ORCID 0000-0003-2510-8993); name matching relies on catalog metadata, not on verified biography.