Alibaba Group researcher in agentic RL for LLM agents and training-free sparse attention for long-context inference.
2
papers here
2
reviewed by Evolve
0
first-author
0
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—
The snapshot lacks publication-count data, so output cannot be scored. This axis would measure total paper volume and per-year production across venues; only two reviewed 2026 arXiv entries are on file so far.
Impact—
No citation counts or h-index are available, so impact is unscored. This axis would measure citation accumulation, per-paper influence, and uptake of the agentic-RL and sparse-attention methods by other research groups.
Focus0/100
The focus score is 0, reflecting a field distribution split evenly, 50/50, between reinforcement learning for LLM agents and efficient long-context LLM inference via sparse attention. The two tracks are technically related — both target LLM capability — but the profile shows a dual-track pattern rather than concentration in a single line of work, so depth in one subfield cannot be asserted from this data.
Collaboration58/100
The collaboration score is 58, based on 17 distinct coauthors in the snapshot. The same core group — Feida Zhu, Pengkun Jiao, and Steven Hoi — recurs across both reviewed papers, indicating work inside a stable industrial team at Alibaba Group rather than a wide academic network. On both 2026 papers the researcher sits in a middle author position without corresponding-author role, consistent with contributor roles in a larger lab group.
Activity—
Recent-activity data is insufficient for a score. This axis would measure publishing cadence and recency, such as outputs per year over recent years; the snapshot only confirms two 2026 preprints.
Reproducibility55/100
Reproducibility scores 55, computed as the mean over 2 assessed papers. Both reviewed papers — ProCredit and the SGDR block-sparse attention work — are rated medium reproducibility readiness. ProCredit reports clean ablations but is verifier-bound in scope; SGDR reports a 5.03x speedup at 128K context, yet thin gains over Prism and unreported variance temper confidence in the numbers.
Research summary
Yiran Zhong is a machine-learning researcher affiliated with Alibaba Group whose current work spans two complementary tracks in large language model systems. The first track is reinforcement learning for LLM agents: the 2026 arXiv paper ProCredit reframes credit assignment by converting verifiable acceptance checks into per-turn progress rewards, yielding consistent gains across three AppWorld scales with clean ablations, though its scope remains bounded by verifier availability. The second track is efficient long-context inference: the 2026 SGDR paper introduces training-free, pre-RoPE block-sparse attention routing built on offline geometric priors, reporting a 5.03x speedup at 128K context, while gains over the Prism baseline are thin and variance is unreported. On both papers Zhong appears in a middle author position without corresponding-author role, working alongside a recurring core team that includes Feida Zhu, Pengkun Jiao, and Steven Hoi among 17 distinct coauthors. The bibliometric field distribution splits evenly, 50/50, between the two tracks, and the reproducibility dimension averages 55 over two medium-readiness papers. Scores for output, impact, and activity are not yet computable from the available snapshot, so this profile — resolved from public bibliographic data under ORCID 0000-0003-1404-3610 — should be read as an early-stage industrial-research profile.
Collaboration pattern
The snapshot records 17 distinct coauthors. Feida Zhu, Pengkun Jiao, and Steven Hoi recur across both reviewed 2026 papers, indicating a stable Alibaba Group research team. Yiran Zhong holds middle author positions without corresponding-author roles, consistent with contributing inside a larger industrial lab group rather than leading an independent academic group. All catalogued collaboration evidence is industrial-lab based.
Research keywords
Highlights
ProCredit: turning outcome rewards into per-turn progress credit
SGDR: training-free block-sparse attention for 128K contexts
A dual-track applied LLM research signature
The bibliometric field distribution splits 50/50 between agentic RL reward design and sparse-attention inference efficiency, and both 2026 papers were built by the same core coauthor team — an applied, production-oriented LLM research pattern rather than a single deep subfield.
Timeline
- 2026The ProCredit preprint appears on arXiv, reframing credit assignment in agentic reinforcement learning as per-turn progress credit for LLM agents.
- 2026The SGDR block-sparse attention preprint appears on arXiv, reporting a 5.03x inference speedup at 128K context with training-free routing.
- 2026Both 2026 papers are built with the recurring core team of Feida Zhu, Pengkun Jiao, and Steven Hoi, anchoring the researcher's current industrial collaboration base within Alibaba Group.
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 scores are derived from OpenAlex-based data; citation counts, h-index, and yearly output are absent from this snapshot.
- The Evolve catalog holds only two reviewed arXiv preprints, both from 2026; no peer-reviewed venue record is included.
- This profile is name-resolved from public bibliographic data linked to ORCID 0000-0003-1404-3610; Yiran Zhong is a common name and identity beyond these records is not asserted.
- Output, impact, and activity dimensions lack underlying data and are described qualitatively rather than scored or estimated.