Zhejiang University researcher advancing agentic RL and LLM post-training methods (RetireOPD, PaperGym, TASPO)
3
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
1
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—
No output score is available: the OpenAlex record lacks reliable work-level totals for this name, so productivity cannot be computed yet. This axis would normally measure total publications, venue spread, and the yearly distribution of output.
Impact—
Impact cannot be scored: citation aggregates are absent from the snapshot, which shows i10-index 0 and h-index 1. This axis would measure citations, adoption of the proposed RL methods, and downstream benchmark usage.
Focus100/100
Focus is maximal at 100/100. The OpenAlex topic profile concentrates in topic modeling (14 works) and NLP techniques (10), with smaller clusters in reinforcement learning in robotics (5) and AI problem solving and planning (5); Evolve's field distribution collapses to a single label — LLM post-training / agentic RL. All three catalog papers pursue the same GRPO-plus-distillation recipe, indicating a deliberately narrow research line.
Collaboration62/100
Collaboration scores 62/100. The record shows 22 distinct co-authors over the recent three-year window, and the frequent-collaborator list is anchored by Zhejiang University colleagues — Weiming Lu, Yuhan Wang, Zhengxi Lu, Yuchen Yan, Kaitao Song, Jun Xiao, Yueting Zhuang, Yongliang Shen and Yan Yu among them. All three Evolve-reviewed papers place this researcher in a middle author position without corresponding status.
Activity91/100
Activity is high at 91/100. The snapshot records 22 works in the recent three-year window, with counts-by-year jumping from 2 works in 2025 to 20 in 2026 — a steep ramp for an early-stage record. All three Evolve-reviewed preprints (RetireOPD, PaperGym, TASPO) are 2026 arXiv postings, consistent with this surge in output.
Reproducibility55/100
Reproducibility scores 55/100, but the estimate rests on a single reviewed paper (reproMean 55 over reproPapers 1), so treat it as a thin signal. Evolve's verdicts flag single-run results with no reported variance (RetireOPD), LLM-judged gains without significance testing (PaperGym), and draft artifacts that lower trust (TASPO); all three papers are rated medium repro-readiness.
Research summary
Wenqi Zhang is a researcher at Zhejiang University whose recent work centers on post-training large language models for agentic settings, particularly reinforcement-learning recipes built around GRPO and distillation. The Evolve catalog holds three 2026 arXiv preprints: RetireOPD, which adds self-retiring on-policy distillation to GRPO agents and reports +12–19 point gains on ALFWorld/WebShop; PaperGym, a rubric-centered framework for research-plan generation trained with OPSD+GRPO on leakage-controlled rubric data (3.7% criterion leakage); and TASPO, which reconciles process supervision with outcome-based credit and averages +10.6% across nine agent-RL settings. The bibliometric snapshot (OpenAlex) shows a steep ramp — 2 works in 2025 rising to 20 in 2026, with 22 works over the recent three-year window — a maximal focus score of 100, and a field distribution that collapses to a single label, LLM post-training / agentic RL. Collaboration is moderate at 62/100 with 22 distinct co-authors, and all three reviewed papers are middle-author contributions. Reproducibility currently reads 55/100 from a single reviewed paper. Citation-based output and impact cannot be scored: the record carries an h-index of 1 and i10-index of 0, so this early-stage profile describes a fast-moving methods contributor whose impact data has not yet caught up with output.
Collaboration pattern
The record shows 22 distinct co-authors in the recent three-year window (collaboration score 62/100). The frequent-collaborator list is anchored by Zhejiang University colleagues including Weiming Lu, Yuhan Wang, Zhengxi Lu, Yuchen Yan, Kaitao Song, Jun Xiao, Yueting Zhuang, Yongliang Shen and Yan Yu. All three Evolve-reviewed papers place Wenqi Zhang in a middle author position without corresponding status, consistent with a contributor inside a larger LLM-agents group.
Research keywords
Highlights
Self-retiring on-policy distillation for agentic RL
Rubric-centered evolution for research-plan generation
Privileged process supervision reconciled with outcome credit
Timeline
- 2025Earliest signal in the snapshot: OpenAlex counts-by-year records 2 works in 2025.
- 2026Output ramps to 20 works in the 2026 count, a steep rise over 2025's 2, with 22 works across the recent three-year window.
- 2026RetireOPD preprint proposes self-retiring on-policy distillation for GRPO agents, reporting +12–19 points on ALFWorld/WebShop.
- 2026Two further preprints appear: PaperGym (rubric-centered research-plan generation, 3.7% leakage) and TASPO (+10.6% average over 9 agent-RL settings).
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 OpenAlex only; the record shows h-index 1 and i10-index 0, and "Wenqi Zhang" is a common name — this profile is name-resolved from public bibliographic data, with no ORCID link provided.
- Output and impact dimensions are unscored because reliable work totals and citation aggregates are missing from the snapshot.
- All three Evolve-reviewed papers are 2026 arXiv preprints rather than peer-reviewed venues, and reproducibility rests on a single reviewed paper.
- Counting windows are narrow: counts-by-year covers only 2025–2026, so longer-term productivity or impact trends cannot be assessed.