ML systems researcher at University of Georgia spanning GPU-cluster training efficiency and on-device LLM inference.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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 axis measures publication volume and productivity over time, but the current snapshot contains only the two 2026 papers reviewed by Evolve, with no aggregate record counts available to estimate output.
Impact—
No score yet: impact would track citations, usage, and downstream adoption of this researcher's work, but citation records are absent from the provided data, so no estimate is offered.
Focus0/100
The focus score is 0, the widest spread in the snapshot: the field distribution splits exactly evenly (0.5/0.5) between mobile/edge machine learning systems and GPU/HPC machine learning systems for deep learning. The two halves are nonetheless adjacent — one targets on-device LLM inference via KV caching on mobile SoCs, the other GPU-cluster JEPA training efficiency — so the profile reads as two angles on efficient ML systems rather than unrelated directions.
Collaboration50/100
Collaboration scores 50 on a moderate base of 11 distinct coauthors in this snapshot. Recurring partners include Wei Niu and Md Musfiqur Rahman Sanim, each appearing on multiple cataloged papers, alongside Zhihao Shu, Jie Hu, Kun Yuan, Minghai Qin, Bahram Afsharmanesh, and Amirali Mirian. Author positions (last on one paper, middle on the other, neither corresponding) suggest senior oversight of group-produced systems work.
Activity—
Not yet scored: activity would measure recent publication tempo and year-over-year research cadence, but the window here holds only two papers, both dated 2026, which is too thin to characterize an activity level.
Reproducibility38/100
Reproducibility readiness averages 38 across exactly two assessed papers. The mask-aware JEPA training paper rates medium, with profiling-driven methodology and ablations but an internal speedup inconsistency and no downstream evaluation; TierKV rates low, with a 17.6x headline benchmarked against MNN-LLM and accuracy dips up to 7 points. Both are arXiv preprints, so code and artifact availability remains the key gap.
Research summary
Gagan Agrawal is affiliated with the University of Georgia and works in machine learning systems, with a catalog profile (name-resolved via ORCID 0000-0002-2609-1428 from public bibliographic data) that currently spans two 2026 arXiv papers reviewed by Evolve. Both sit at the intersection of efficient deep learning and computer systems: "Mask-Aware Execution for Efficient JEPA Training" targets GPU-cluster training efficiency for self-supervised JEPA models, while "TierKV" addresses long-context LLM inference on mobile SoCs through multi-tier KV caching that mixes exact, SVD-based, and flash-backed tiers. The author-position pattern is senior-leaning — last author on the JEPA paper and a middle author on TierKV, neither marked corresponding — consistent with contributing through students and collaborators rather than hands-on first-author execution. Evolve's review notes a claimed 1.7x JEPA training speedup from execution restructuring alone, profiling-driven and ablated but carrying an internal speedup inconsistency and no downstream evaluation, and a 17.6x TierKV headline measured against MNN-LLM, with accuracy dips reaching 7 points. The field distribution in this snapshot splits evenly (0.5/0.5) between mobile/edge ML systems and GPU/HPC systems for deep learning; 11 distinct coauthors appear, with Wei Niu and Md Musfiqur Rahman Sanim recurring. Reproducibility readiness averages 38 across the two papers, rated medium for one and low for the other. Output, impact, and activity axes lack sufficient data and are left unscored.
Collaboration pattern
Within this snapshot Gagan Agrawal works with 11 distinct coauthors across the two cataloged systems papers. The most recurring names are Wei Niu and Md Musfiqur Rahman Sanim, each appearing more than once; other collaborators include Zhihao Shu, Jie Hu, Kun Yuan, Minghai Qin, Bahram Afsharmanesh, and Amirali Mirian. Taking the last-author slot on the JEPA training paper and a middle slot on TierKV (neither corresponding), the pattern fits a senior collaborator supervising group-built systems projects.
Research keywords
Highlights
Mask-aware execution restructures JEPA training, claiming 1.7x GPU speedup
TierKV enables long-context on-device LLMs with predictive multi-tier KV caching
Timeline
- 2026First cataloged paper "Mask-Aware Execution for Efficient JEPA Training" appears on arXiv, with Gagan Agrawal in the last author position.
- 2026Second cataloged paper "TierKV: Long-Context On-Device LLMs via Predictive Multi-Tier KV Caching" appears on arXiv, with Gagan Agrawal as a middle author.
- 2026Both papers enter Evolve's reviewed catalog, establishing an early full-tier profile spanning GPU-cluster training efficiency and on-device LLM inference.
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.
- Radar scores and bibliometrics are derived from OpenAlex-based public bibliometric data only; the output, impact, and activity axes lack sufficient data to score.
- The reviewed catalog covers only two 2026 arXiv preprints; no peer-reviewed venue records are included in this snapshot.
- Identity is name-resolved via ORCID 0000-0002-2609-1428 from public bibliographic data; "Gagan Agrawal" is a shared name, and disambiguation beyond the ORCID link is not guaranteed.
- The reproducibility mean of 38 rests on exactly two assessed papers and may shift as catalog coverage grows.