Alibaba ML researcher: agentic RL progress credit for LLM agents and sparse attention for long-context inference
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-09) 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—
This axis would measure publication volume and productivity over time, but the linked OpenAlex snapshot contains only the two 2026 arXiv entries, so per-year output counts are insufficient and no output score can be responsibly computed.
Impact—
This axis would measure citation-based influence such as total citations, h-index and i10-index. The snapshot reports zeros that reflect an incomplete bibliographic record rather than measured influence, so the underlying data is insufficient for scoring.
Focus0/100
The focus score is 0, reflecting topical spread rather than concentration: the field distribution splits evenly (0.5/0.5) between reinforcement learning for LLM agents and efficient long-context LLM inference, and each of the five tagged topics — topic modeling, advanced neural network applications, NLP techniques, speech and audio processing, and hand gesture recognition — appears only once, so no dominant single research line emerges from the current record.
Collaboration58/100
The collaboration score is 58, based on 17 distinct coauthors across the two 2026 papers. Feida Zhu and Pengkun Jiao recur in the frequent-collaborator list, alongside Yiran Zhong, Ming Ma, Yi Zhu, Chonghan Liu, Qichao Wang, Yanhao Jia and Tianming Yang. Authorship positions — last author on ProCredit, middle author on SGDR — suggest a senior contributor within team-based projects rather than a solo researcher.
Activity32/100
The activity score is 32. countsByYear records a single year, 2026, with 2 works, and recent3yWorks is also 2 — both tied to the two arXiv preprints in this profile. With no earlier publications captured in the snapshot, measured recent activity is low and concentrated entirely within one year.
Reproducibility55/100
The reproducibility score is 55, averaged over 2 reviewed papers, each rated medium reproducibility readiness. ProCredit shows consistent gains across three AppWorld scales with clean ablations, though its scope is bound to available verifiers; SGDR delivers a 5.03x speedup at 128K context but reports thin margins over Prism and no variance, both of which temper replication confidence.
Research summary
Steven Hoi is a machine-learning researcher affiliated with Alibaba Group (China). His current public record, as captured by OpenAlex, centers on two 2026 arXiv preprints that split evenly between two directions: reinforcement learning for LLM agents and efficient long-context LLM inference via sparse attention. On ProCredit, where he is the last author, the team converts verifiable acceptance checks into per-turn progress rewards for agentic reinforcement learning, reporting consistent gains across three AppWorld scales with clean ablations while noting the method stays bound to available verifiers. On SGDR (Block-Sparse Attention with Semantic-Geometric Decoupled Routing), a middle-author paper, the group introduces training-free pre-RoPE block routing with offline geometric priors, achieving a 5.03x speedup at 128K context, though margins over Prism are thin and variance is unreported. Bibliometrically, the snapshot lists 17 distinct coauthors (collaboration score 58), a reproducibility score of 55 across 2 reviewed papers with medium readiness, and an activity score of 32, since countsByYear contains only 2026 with 2 works. Reported h-index and i10-index are 0 in this record, and the output and impact dimensions lack sufficient data. The profile is name-resolved from public bibliographic data; no ORCID is linked.
Collaboration pattern
The record lists 17 distinct coauthors across the two 2026 arXiv papers, with Feida Zhu and Pengkun Jiao recurring in the frequent-collaborator list alongside Yiran Zhong, Ming Ma, Yi Zhu, Chonghan Liu, Qichao Wang, Yanhao Jia and Tianming Yang. Sitting last author on ProCredit and middle author on SGDR, Steven Hoi appears to operate as a senior collaborator inside larger Alibaba team projects rather than as a solo first author.
Research keywords
Highlights
Progress-credit rewards for agentic RL
Training-free block-sparse attention for long contexts
Timeline
- 2026Both papers in the current record appear: countsByYear shows 2 works in 2026 (recent3yWorks also 2), splitting evenly between agentic RL for LLM agents and sparse-attention long-context inference.
- 2026ProCredit: From Outcome Rewards to Progress Credit in Agentic Reinforcement Learning is posted to arXiv, with Steven Hoi as last author; it reports consistent gains across three AppWorld scales.
- 2026Block-Sparse Attention with Semantic-Geometric Decoupled Routing (SGDR) is posted to arXiv with a middle-author position; it reports a 5.03x speedup at 128K context.
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 data comes from OpenAlex only; the profile is name-resolved from public bibliographic data and no ORCID is linked, so identity confidence is moderate.
- The snapshot appears to capture only the two 2026 arXiv preprints; the reported h-index (0) and i10-index (0) reflect this limited record, not verified career-level citation metrics.
- Output and impact dimensions are marked missing because citation and per-year output data are insufficient in the current snapshot.
- Both catalog papers are arXiv preprints with medium reproducibility readiness; peer-review status is outside the provided data.