Feida Zhu

Alibaba Group (Cayman Islands)

Unclaimed · AI-structured profileOpenAlex

Agentic RL researcher at Alibaba Group shaping reward design and credit assignment for LLM-based mobile agents

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

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—

Output is not scored yet: the OpenAlex record behind this profile indexes only 2 recent works, so total publication volume and productivity cannot be measured reliably. This axis would count papers per year across the full indexed publication history.

Impact—

Impact is not scored yet: citation data for this record are unreliable — the h-index reads 0 despite active 2026 preprints. This axis would measure citations, h-index, and influence across the indexed literature.

Focus0/100

The focus score of 0 reflects even dispersion in this record: the OpenAlex field distribution splits 0.5/0.5 between 'machine learning — reinforcement learning for LLM agents' and 'machine learning — LLM-based agents', and topic tags spread across five areas (reinforcement learning in robotics ×2; robot manipulation, planning, games, and XAI ×1 each). The two reviewed 2026 papers actually share one tight reward-design agenda, so part of this dispersion looks like a field-labeling artifact.

Collaboration71/100

Collaboration scores 71, grounded in 35 distinct coauthors within the recent OpenAlex window. Pengkun Jiao and Steven Hoi are the only recurring names in the frequent-coauthor list, while Ming Ma, Yi Zhu, Yiran Zhong, Chonghan Liu, Qichao Wang, Yanhao Jia, and Tianming Yang appear alongside them. On both cataloged 2026 papers Zhu holds a middle author position without corresponding authorship, fitting a contributing role within larger Alibaba research teams.

Activity32/100

Activity scores 32: recent3yWorks is 2 and countsByYear shows a single year, 2026, with 2 outputs — both arXiv preprints. All indexed output is compressed into one year with no earlier record, so the low score reflects a thin recent window rather than dormancy; the OpenAlex profile may simply undercount work from before 2026.

Reproducibility38/100

Reproducibility averages 38/100 across the 2 reviewed papers (reproPapers: 2). ProCredit rates medium: consistent gains across three AppWorld scales with clean ablations, though bounded by verifier availability. Qwen-Planner-Agent rates low: the headline 77.05% on MobilePA-Bench comes from a self-built benchmark without significance tests. Both papers are arXiv preprints with Zhu in a middle, non-corresponding position; the pair's mean is dragged down by the weaker of the two.

Research summary

Feida Zhu is a researcher affiliated with Alibaba Group whose recent published work centers on reinforcement learning for large language model agents. The Evolve catalog contains two 2026 arXiv papers, on both of which Zhu is a middle-position author rather than corresponding author. The first, Qwen-Planner-Agent, presents a closed-loop AI-for-AI framework that trains a 27B-parameter mobile planner reaching 77.05% on the self-built MobilePA-Bench, though the reported gains lack statistical significance testing. The second, ProCredit, converts verifiable acceptance checks into per-turn progress rewards for agentic reinforcement learning, delivering consistent gains across three AppWorld scales with clean ablations while remaining verifier-bound. Together these define a narrow but coherent agenda — reward shaping and credit assignment for agentic RL — mirrored in the OpenAlex field distribution, which splits evenly at 0.5 for LLM-based agents and 0.5 for reinforcement learning for LLM agents. The record lists 35 distinct coauthors in the recent window, with Pengkun Jiao and Steven Hoi recurring among frequent collaborators, and topic tags spanning robotics RL, planning, games, and explainable AI. Activity is concentrated — recent3yWorks is 2, both outputs dated 2026 — and reproducibility readiness averages 38/100 across the pair, rated low for the planner work and medium for ProCredit.

Collaboration pattern

Zhu's recorded pattern is team-based: 35 distinct coauthors appear in the recent OpenAlex window, with Pengkun Jiao and Steven Hoi the only recurring names in the frequent-coauthor list. On both 2026 catalog papers Zhu is a middle-position, non-corresponding author, consistent with specialist contributions (reward design, agent harnessing) inside larger Alibaba project teams; other listed collaborators include Ming Ma, Yi Zhu, Yiran Zhong, Chonghan Liu, Qichao Wang, Yanhao Jia, and Tianming Yang.

Research keywords

LLM agentsAgentic reinforcement learningCredit assignmentVerifiable rewardsReward shapingGRPOMobile agentsAgent harnessClosed-loop AI-for-AI frameworksBenchmark evaluation

Highlights

Closed-loop AI-for-AI training lifts a 27B mobile planner to 77.05% on MobilePA-Bench
The Qwen-Planner-Agent pipeline couples a mobile agent harness with agentic reinforcement learning (GRPO, reward shaping) to train a 27B planner that tops the self-built MobilePA-Bench at 77.05%. Evolve flags low reproducibility readiness: the gains lack significance tests and the benchmark is not community-standard.
Progress credit turns outcome rewards into per-turn signal for agentic RL
ProCredit converts verifiable acceptance checks into per-turn progress rewards, yielding consistent gains across three AppWorld scales with clean ablations. Evolve rates its reproducibility readiness medium, noting the method's scope is bounded by the availability of a reliable verifier.
One reward-design agenda across two 2026 papers
Both cataloged papers — one on mobile planning, one on credit assignment — rely on GRPO and reward shaping for LLM agents, matching the OpenAlex field split of 0.5/0.5. Zhu appears as a middle-position, non-corresponding author on both, suggesting a shared team effort at Alibaba rather than two independent lines.

Timeline

2026Qwen-Planner-Agent released on arXiv: a closed-loop AI-for-AI framework trains a 27B mobile planner reaching 77.05% on the self-built MobilePA-Bench (reproducibility readiness: low).
2026ProCredit released on arXiv: verifiable acceptance checks are turned into per-turn progress rewards, with consistent gains across three AppWorld scales (reproducibility readiness: medium).
2026Both papers enter the Evolve catalog under the LLM-agent field split (0.5/0.5), with 35 distinct coauthors recorded in the recent OpenAlex window.

Papers in the Evolve catalog

Frequent co-authors

Pengkun Jiao×2 Steven Hoi×2 Ming Ma×1 Yi Zhu×1 Yiran Zhong×1 Feida Zhu×1 Chonghan Liu×1 Pengkun Jiao×1 Qichao Wang×1 Yanhao Jia×1 Tianming Yang×1 Steven Hoi×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 scores come from a single OpenAlex record (identity: openalex) and may undercount this person's full publication history.
No ORCID or country is available; the profile is name-resolved from public bibliographic data and homonymous authors cannot be ruled out.
Output and impact axes are unscored because citation counts and a complete publication list are not reliably indexed (the recorded h-index is 0).
Reproducibility and quality assessments cover only the 2 papers reviewed in the Evolve catalog.