Feida Zhu

Alibaba Group (China) · Cayman Islands

Unclaimed · AI-structured profileOpenAlex

Machine learning researcher at Alibaba Group working on RL for LLM agents and efficient long-context LLM inference

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-09) 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 volume cannot be scored from current data: the snapshot contains no publication-count series beyond the 2 works recorded in 2026. This axis would measure total and per-year research productivity, so it is left unscored rather than estimated.

Impact—

Impact is unscored: no citation counts are available in the snapshot, and the reported h-index and i10-index are 0 because both 2026 preprints are too recent to accrue citations. This axis would measure citation influence and community uptake.

Focus0/100

The focus axis scores 0, mirroring an exactly even field distribution: 0.5 for reinforcement learning for LLM agents and 0.5 for efficient long-context LLM inference via sparse attention. The five OpenAlex topics attached to this identity (Topic Modeling, Advanced Neural Network Applications, Natural Language Processing Techniques, AI-based Problem Solving and Planning, Multimodal Machine Learning Applications, each with count 1) likewise spread across adjacent ML areas with no dominant concentration.

Collaboration58/100

Collaboration scores 58, a mid-range value. The two catalog papers together involve 17 distinct coauthors, and the frequent-coauthor list is headed by Pengkun Jiao and Steven Hoi, who each recur across the record. On both papers Zhu sits in a middle author position without corresponding-author role, consistent with contributing within larger, industry-scale teams rather than leading a small group.

Activity32/100

Activity scores 32, on the low side. The countsByYear series contains a single entry — 2 works in 2026 — and recent3yWorks is likewise 2, meaning this OpenAlex identity shows no recorded output in the earlier years of the three-year window. Such a thin series bounds how much recent-productivity signal the axis can carry.

Reproducibility55/100

Reproducibility scores 55, a reproMean computed over 2 papers, both rated medium readiness by Evolve. ProCredit is credited with clean ablations that support replication, while SGDR reports a 5.03x speedup at 128K but thin gains over Prism with variance unreported — the kind of missing statistics that holds its readiness rating down.

Research summary

Feida Zhu is a machine learning researcher affiliated with Alibaba Group (China), profiled here from an OpenAlex bibliographic snapshot and the two 2026 arXiv preprints reviewed in the Evolve catalog. The reviewed output splits evenly across two subfields — reinforcement learning for LLM agents and efficient long-context LLM inference via sparse attention — each accounting for 0.5 of the field distribution, which is consistent with the focus score of 0. The catalog entries are ProCredit, which converts verifiable acceptance checks into per-turn progress rewards for agentic RL with consistent gains across three AppWorld scales, and SGDR, a training-free block-sparse attention router reporting a 5.03x speedup at 128K context. On both papers Zhu sits in a middle author position without corresponding-author role, within teams drawing on 17 distinct coauthors. Activity scores 32, with countsByYear recording 2 works in 2026 and recent3yWorks also 2. Reproducibility scores 55, a reproMean over 2 papers both rated medium readiness. Citation-based impact is not yet measurable: the snapshot reports h-index and i10-index of 0, and the output and impact axes lack sufficient data to score.

Collaboration pattern

Collaboration scores 58 with 17 distinct coauthors across the two catalog papers. Pengkun Jiao and Steven Hoi are the only recurring names in the frequent-coauthor list (each appearing twice), alongside Yiran Zhong, Ming Ma, Yi Zhu, Chonghan Liu, Qichao Wang, Yanhao Jia and Tianming Yang. Both papers place Zhu in a middle, non-corresponding position, suggesting a contributor role within larger industry teams rather than senior or sole leadership.

Research keywords

agentic reinforcement learningcredit assignmentverifiable rewardsreward shapingGRPOLLM agentsblock-sparse attentiontraining-free routingRoPElong-context inferenceoffline calibration

Highlights

Progress credit for agentic RL: ProCredit turns acceptance checks into per-turn rewards
In this 2026 arXiv paper Zhu contributes as a middle author to a method that converts verifiable acceptance checks into per-turn progress rewards for agentic reinforcement learning, moving beyond sparse outcome rewards. Evolve notes consistent gains across three AppWorld scales and clean ablations, while flagging the method's verifier-bound scope.
Training-free long-context speedup: SGDR block-sparse attention at 5.03x @128K
The second 2026 preprint, again with Zhu as a middle author, introduces Semantic-Geometric Decoupled Routing for block-sparse attention: training-free pre-RoPE block routing built on offline geometric priors, delivering a 5.03x speedup at 128K context. Evolve judges the gains over Prism thin and notes variance is unreported.
Medium reproducibility readiness across both 2026 releases
Evolve rates both catalog papers medium on reproducibility readiness, yielding this profile's reproMean of 55 over 2 papers. ProCredit's clean ablations support replication, while SGDR's unreported variance is the main gap for anyone re-implementing it.

Timeline

2026ProCredit preprint posted to arXiv, applying per-turn progress-credit rewards to agentic reinforcement learning for LLM agents, with consistent gains across three AppWorld scales.
2026SGDR block-sparse attention preprint posted to arXiv, reporting a training-free routing scheme with a 5.03x inference speedup at 128K context.
2026OpenAlex countsByYear records 2 works for this identity in 2026, with recent3yWorks also at 2 — the only years with recorded output.

Papers in the Evolve catalog

Frequent co-authors

Yiran Zhong×2 Pengkun Jiao×2 Steven Hoi×2 Ming Ma×1 Yi Zhu×1 Chonghan Liu×1 Qichao Wang×1 Yanhao Jia×1 Tianming Yang×1 Feida Zhu×1 Pengkun Jiao×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.

Bibliometrics come from an OpenAlex snapshot (identity: openalex); 'Feida Zhu' is name-resolved from public bibliographic data and may overlap with other researchers of the same name. No ORCID is provided to disambiguate.
The output and impact axes are unscored: no citation counts are available and the snapshot reports h-index and i10-index of 0, consistent with two very recent 2026 preprints.
Scope is limited to the two 2026 arXiv papers reviewed in the Evolve catalog; any earlier work behind the five recorded OpenAlex topics is not covered in this profile.
Affiliation and country are recorded as provided ('Alibaba Group (China)', country code 'KY') and have not been independently verified.