ML systems researcher spanning mobile/edge LLM inference and GPU/HPC training efficiency at University of Georgia.
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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: this snapshot lacks the publication-volume data this axis measures — total outputs, venue mix, and yearly production. With only two arXiv preprints reviewed, no productivity estimate is made here.
Impact—
No citation record is available in this snapshot, so impact — citations, downstream adoption, and influence on later systems work — cannot be scored. Both papers are recent 2026 preprints, too early for meaningful citation signals.
Focus0/100
The focus axis scores 0 in this snapshot because the reviewed output splits exactly 50/50 between two field labels: mobile/edge machine-learning systems and GPU/HPC machine-learning systems. Read literally, that is no single-domain concentration; read practically, the two halves are adjacent branches of one efficiency agenda — KV-cache and on-device inference on one side, kernel-level training restructuring on the other — so the spread reflects a systems stack rather than scattered interests.
Collaboration50/100
Collaboration scores 50 (mid-range) on 11 distinct coauthors recorded for this researcher. The frequent-coauthor list shows repeats — Gagan Agrawal, Wei Niu, and Md Musfiqur Rahman Sanim each appear more than once — alongside Jie Hu, Kun Yuan, Minghai Qin, Bahram Afsharmanesh, and Amirali Mirian. The pattern is team-based systems research: TierKV was first-authored within a multi-person group, and the JEPA training paper added middle-authorship work in a larger team.
Activity—
Insufficient data: this axis would measure research tempo — yearly output cadence, submission continuity, and recent engagement — but the snapshot only fixes both reviewed papers in 2026 and carries no multi-year production record, so no activity level is scored.
Reproducibility38/100
Reproducibility scores 38, the mean readiness Evolve assigned across exactly two reviewed papers. The spread is informative: the JEPA paper rates medium — profiling-driven with ablations, but carrying an internal speedup inconsistency and no downstream evaluation — while first-authored TierKV rates low, with its 17.6× headline measured against MNN-LLM and accuracy dips reaching 7 points. Neither paper yet meets a high replication bar.
Research summary
Zhihao Shu is a machine-learning-systems researcher at the University of Georgia whose reviewed work to date spans the two ends of the deep-learning efficiency problem: datacenter-scale training and on-device inference. The Evolve catalog contains two 2026 arXiv papers. The first, TierKV (first-authored), tackles long-context LLM inference on mobile SoCs through a predictive solver that tiers the KV cache across exact, SVD-compressed, and flash-offloaded storage; Evolve judges it credible systems work while flagging that the 17.6× headline is measured against MNN-LLM and that accuracy dips reach 7 points. The second, Mask-Aware Execution for Efficient JEPA Training (middle authorship), reports a 1.7× training speedup from execution restructuring alone — kernel fusion, sparse computation, and token routing — with profiling-driven analysis and ablations, though Evolve notes an internal speedup inconsistency and the absence of downstream-task evaluation. Bibliometrically, the profile shows 11 distinct coauthors and a field distribution split evenly (50/50) between mobile/edge ML systems and GPU/HPC machine-learning systems; collaboration scores 50 and reproducibility 38, the mean over two papers. Output, impact, and activity metrics are not yet available. This is an early-stage, arXiv-centric profile: peer-review outcomes, citation record, and career history remain outside the current data scope.
Collaboration pattern
Collaboration sits mid-range (score 50) on 11 distinct coauthors. The frequent-coauthor list repeats Gagan Agrawal, Wei Niu, and Md Musfiqur Rahman Sanim, and also includes Jie Hu, Kun Yuan, Minghai Qin, Bahram Afsharmanesh, and Amirali Mirian. The 2026 pattern is team-based systems work: first-authorship on TierKV within a multi-person group and middle-authorship on the JEPA training paper, consistent with lab-driven, multi-contributor engineering projects rather than solo output.
Research keywords
Highlights
TierKV: predictive multi-tier KV caching for long-context on-device LLMs
Mask-aware execution claims 1.7× faster JEPA training
A footprint spanning edge inference and datacenter training
Across the two reviewed papers, field distribution splits 50/50 between mobile/edge ML systems and GPU/HPC ML systems, with 11 distinct coauthors — positioning the researcher on both sides of the deep-learning efficiency problem in 2026.
Timeline
- 2026First-author arXiv posting of TierKV, introducing predictive multi-tier KV caching for long-context on-device LLMs.
- 2026Co-authored arXiv posting of Mask-Aware Execution for Efficient JEPA Training, reporting a 1.7× training speedup via execution restructuring alone.
- 2026Bibliometric footprint for the year consolidates: 11 distinct coauthors and an even 50/50 field split between mobile/edge ML systems and GPU/HPC ML systems.
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.
- Bibliometrics are drawn from OpenAlex-based public bibliographic data only; citation counts, venue-level output, and h-index were not part of this snapshot.
- Output, impact, and activity axes are unscored because the underlying data was insufficient; no estimates were substituted.
- Both reviewed papers are 2026 arXiv preprints; peer-review status, published venues, and post-review revisions are outside the current data.
- Profile is name-resolved from ORCID 0009-0008-8749-7115 and public bibliographic records; scope is limited to the provided snapshot and Evolve-reviewed papers.