Machine learning researcher at Alibaba Group working on agentic reinforcement learning and LLM-based mobile agents.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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—
The output axis is not scored yet: the snapshot's h-index and i10-index both read 0, which points to incomplete author disambiguation rather than measured productivity. It would count total publications and their per-year distribution for this identity.
Impact—
Impact is likewise unscored: no citation counts are available for this OpenAlex identity in the snapshot. This axis would measure citation-based influence — total citations, h-index trajectory, and per-paper citation distribution — and stays unscored here.
Focus0/100
Focus scores 0 in this snapshot because the field distribution splits exactly 50/50 between two adjacent subfields — reinforcement learning for LLM agents and LLM-based agents — while topic signals spread thinly across five topics (reinforcement learning in robotics at count 2, plus robot manipulation and learning, planning, games, and explainable AI at 1 each). In practice the two halves are closely related, so the low focus score reads as breadth within one agentic-RL program rather than scattered interests.
Collaboration71/100
Collaboration scores 71, backed by 35 distinct co-authors recorded for this identity in the snapshot window. The recurring names in the frequent-coauthor list are Feida Zhu and Pengkun Jiao (each appearing twice), alongside Ming Ma, Yi Zhu, Yiran Zhong, Chonghan Liu, Qichao Wang, Yanhao Jia, and Tianming Yang. Both reviewed 2026 papers place him in the last author position without corresponding-author flags, consistent with a senior, supervisory role in group-produced agentic-RL work.
Activity32/100
Activity scores 32 on a thin recent record: the snapshot counts only 2 works in the recent three-year window, and the per-year counts contain a single entry, 2026, with 2 publications. Both are arXiv preprints reviewed by Evolve that year — ProCredit and Qwen-Planner-Agent — so the visible activity is concentrated in one burst; no earlier years appear in the countsByYear data, limiting what this axis can say.
Reproducibility38/100
Reproducibility averages 38 across the 2 papers Evolve has reviewed. ProCredit is rated medium — consistent gains across three scales in AppWorld with clean ablations, though verifier-bound in scope — while Qwen-Planner-Agent is rated low: its headline result, 77.05% on MobilePA-Bench, comes from a self-built benchmark without significance tests. The pattern suggests strong empirical pipelines paired with evaluation practices that would need external benchmarks and statistical rigor to replicate confidently.
Research summary
Steven Hoi is profiled here as a machine learning researcher affiliated with Alibaba Group, with identity resolved from OpenAlex bibliographic data. The Evolve catalog reviews two of his 2026 arXiv preprints, both in the space of LLM-based agents and agentic reinforcement learning, with him listed in the last author position — the slot conventionally reserved for senior or supervising authors — though neither flags him as corresponding author. The first, Qwen-Planner-Agent, presents a closed-loop AI-for-AI framework that trains a 27B mobile planner reported to top MobilePA-Bench at 77.05%, with the caveat that gains lack significance tests on a self-built benchmark. The second, ProCredit, converts verifiable acceptance checks into per-turn progress rewards, showing consistent gains across three scales in AppWorld with clean ablations, within a verifier-bound scope. Radar scores place collaboration at 71 (35 distinct co-authors in the snapshot window), reproducibility at 38 (the mean across the two reviewed papers), and activity at 32 (2 works in the recent three-year window), while output and impact remain unscored for lack of citation-level data. Topic signals cluster around reinforcement learning in robotics, robot manipulation and learning, planning, games, and explainable AI.
Collaboration pattern
Steven Hoi's collaboration footprint in this snapshot is broad: 35 distinct co-authors, with Feida Zhu and Pengkun Jiao the only names appearing more than once in the frequent-coauthor list, alongside Ming Ma, Yi Zhu, Yiran Zhong, Chonghan Liu, Qichao Wang, Yanhao Jia, and Tianming Yang. On both reviewed 2026 papers he sits in the last author position without the corresponding-author flag — a pattern consistent with senior supervision of group-built agentic-RL projects rather than hands-on first-author work.
Research keywords
Highlights
From outcome rewards to per-turn progress credit
Closed-loop AI-for-AI training for mobile planner agents
Timeline
- 2026ProCredit preprint appears on arXiv, reframing outcome rewards as per-turn progress credit for agentic reinforcement learning; Hoi is listed as last author.
- 2026Qwen-Planner-Agent preprint appears on arXiv, introducing a closed-loop AI-for-AI framework for mobile planner agents; Hoi is again last author.
- 2026Both preprints enter the Evolve review catalog, forming the current evidence base for this profile with reproducibility ratings of medium and low respectively.
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 snapshot's h-index and i10-index read 0, likely an author-disambiguation artifact, so output and impact are unscored.
- "Steven Hoi" is a name shared by multiple machine-learning researchers; this profile is name-resolved from public OpenAlex bibliographic data, with no ORCID to confirm identity.
- Highlights, timeline, and reproducibility rest on just 2 Evolve-reviewed papers, both 2026 arXiv preprints; earlier output falls outside this snapshot.
- Affiliation is recorded as "Alibaba Group (Cayman Islands)" with country null; institutional detail is taken as provided and not independently verified.