Peking University researcher in efficient LLM inference and RL environments for LLM coding agents
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) 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—
Not yet scored. Registry citation metrics (h-index and i10-index both recorded as 0) are insufficient to characterize this researcher's publication output. This axis would measure total works, annual output volume, and venue tiers once a fuller bibliometric record becomes available.
Impact—
Not yet scored. Citation counts and h-index-based influence data are not available in the current registry (h-index and i10-index recorded as 0). This axis would measure citations, citation trajectory over time, and downstream influence on follow-up work.
Focus0/100
Scores 0/100, reflecting wide topical dispersion in the underlying data: the field distribution splits evenly (50/50) between efficient LLM inference architecture design and machine learning for LLM coding agents, and the topic profile lists five distinct topics with one work each. The two catalog papers address unrelated problems (sparse attention versus RL environment synthesis), so no single research line yet dominates this early profile.
Collaboration71/100
Scores 71/100. The OpenAlex record lists 34 distinct coauthors, and 12 collaborators recur across the profile, including Liang Zhao, Fuli Luo, Jianyu Wei, Yizhao Gao, and Yu Cheng. Both catalog papers place this researcher in a middle author position rather than first or corresponding, consistent with team-based work inside larger groups spanning systems and machine learning rather than a small two-person partnership.
Activity40/100
Scores 40/100. The counts-by-year record shows 3 works indexed in 2026, matching the recent-3-year count of 3, so all visible output is concentrated in a single recent year. Both catalog papers are 2026 arXiv preprints (HySparse2 and CodeMidas) with no earlier indexed publications, indicating an active but very recent start whose sustained cadence is not yet established.
Reproducibility20/100
Scores 20/100, computed from Evolve's review of 2 papers with a mean repro-readiness of 20 (reproMean 20 across reproPapers 2); both catalog entries are rated low. HySparse2's MQA switching confounds its attention comparisons, and CodeMidas reports a single run (5,545 synthesized RL environments lifting MiMo-V2.5 on five benchmarks) without significance tests, so independent replication is not yet straightforward for either paper.
Research summary
Hailin Zhang is a researcher at Peking University whose profile is name-resolved from OpenAlex public bibliographic data. The visible record is early-stage and concentrated in 2026, with 3 works indexed that year and a recent-3-year count of 3. The profile splits evenly between two fields: efficient large language model inference (architecture design) and machine learning for LLM coding agents. Two reviewed arXiv preprints anchor it. HySparse2 (2026) proposes hybrid sparse attention with two-level KV sharing that cuts prefill cost and KV cache while improving retrieval, though Evolve notes an MQA switch that confounds comparisons. CodeMidas (2026) claims that source code alone can yield 5,545 reinforcement learning environments that lift MiMo-V2.5 on five benchmarks, with gains judged plausible but from a single run without significance tests. On both papers the researcher holds a middle author position, not first or corresponding. Collaboration is the strongest measurable axis (71/100): 34 distinct coauthors and 12 recurring collaborators. Focus scores 0/100 given the even field split and five single-work topics, and reproducibility scores 20/100 because both papers are rated low readiness. Output and impact remain unscored: the registry records h-index and i10-index as 0, so citation-based influence cannot yet be characterized.
Collaboration pattern
Collaboration scores 71/100. The OpenAlex record lists 34 distinct coauthors, and 12 recur often, including Liang Zhao, Fuli Luo, Jianyu Wei, Yizhao Gao, Yu Cheng, Yifan Song, and Zihan Jiang. On both catalog papers the researcher sits in a middle author position without corresponding status, suggesting contributor roles within larger Peking University teams spanning systems engineering and machine learning rather than solo-led projects.
Research keywords
Highlights
HySparse2: two-level KV sharing sparse attention for long-context inference
CodeMidas: 5,545 RL coding environments synthesized from source code
Timeline
- 2026HySparse2 preprint released on arXiv: hybrid sparse attention with two-level KV sharing for long-context LLM inference, with middle-author position and low repro-readiness rating.
- 2026CodeMidas preprint released on arXiv: 5,545 agentic coding RL environments synthesized from code itself, lifting MiMo-V2.5 on five benchmarks, again as a middle author.
- 2026OpenAlex record shows 3 works indexed in 2026 (recent-3-year count: 3), with 34 distinct coauthors and a collaboration score of 71/100.
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; identity is name-resolved from public bibliographic records with no linked ORCID, so common-name ambiguity cannot be fully excluded.
- Output and impact axes are unscored: the registry records h-index and i10-index as 0, and citation counts are not yet available.
- Both reviewed papers are 2026 arXiv preprints with middle authorship; peer-reviewed venue status is not reflected in this data.
- The reproducibility score (mean 20) rests on only 2 reviewed papers, both rated low readiness, so it may not generalize.