Machine-learning systems researcher: GPU-efficient JEPA training and multi-tier KV caching for on-device LLMs
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: the provided snapshot lists no publication-count or venue totals, only the two cataloged 2026 preprints. This axis would measure cumulative and yearly paper volume across peer-reviewed venues once a fuller bibliographic record is linked.
Impact—
Not yet scored: no citation counts, h-index, or venue-level influence figures appear in the input data. This axis would measure citations accruing to the two 2026 preprints and their academic uptake once citation tracking becomes available.
Focus0/100
The field distribution in the data splits exactly 50/50 between 'computer systems — mobile/edge machine learning systems' and 'machine learning systems (GPU/HPC for deep learning)', yielding a focus score of 0. The two halves are closely related in practice: one paper optimizes GPU-side JEPA training, the other on-device LLM serving, so the low score reflects breadth across adjacent systems subfields rather than scattered interests.
Collaboration50/100
The collaboration score is 50, based on 11 distinct co-authors in the snapshot. The frequent-coauthor list includes Gagan Agrawal and Wei Niu, each appearing twice, alongside Zhihao Shu, Jie Hu, Kun Yuan, Minghai Qin, Bahram Afsharmanesh, and Amirali Mirian. Both cataloged papers are multi-author systems projects, with Sanim first author on one and a middle author on the other, indicating active team-based work.
Activity—
Not yet scored: the data provides paper years but no per-year publication cadence, so recent research activity cannot be computed. This axis would measure publishing frequency and continuity across the 2026 window and beyond as more records accumulate.
Reproducibility38/100
Reproducibility scores 38 (reproMean across 2 papers). Evolve rates the JEPA training paper 'medium' — profiling-driven and ablated, but with an internal speedup inconsistency and no downstream evaluation — and TierKV 'low', where accuracy dips reach 7 points and the 17.6x speedup is specific to the MNN-LLM baseline. Reproduction attempts should start with the medium-readiness JEPA work.
Research summary
Md Musfiqur Rahman Sanim is an early-career machine-learning systems researcher whose public bibliographic record (ORCID 0009-0009-8431-3699) currently centers on two 2026 arXiv preprints reviewed in the Evolve catalog. The work spans the two halves of an even 50/50 field distribution in the underlying data: GPU/HPC-oriented training efficiency for deep learning, and mobile/edge machine-learning systems for on-device LLM inference. As first author of 'Mask-Aware Execution for Efficient JEPA Training', Sanim proposes restructuring execution for self-supervised JEPA training, claiming a 1.7x speedup from execution restructuring alone; Evolve's review credits the profiling-driven design and ablations while flagging an internal speedup inconsistency and the absence of downstream-task evaluation. As a middle author on 'TierKV', the researcher contributes to predictive multi-tier KV caching for long-context on-device LLMs, where the 17.6x headline speedup is measured against MNN-LLM and accuracy dips reach 7 points. Collaboration data lists 11 distinct co-authors, with Gagan Agrawal and Wei Niu the most recurrent names. Output, impact, and activity metrics are not yet available in the provided snapshot, so this remains an early-stage profile.
Collaboration pattern
The snapshot lists 11 distinct co-authors. The most recurrent names are Gagan Agrawal and Wei Niu, each appearing twice in the frequent-coauthor list, alongside Zhihao Shu, Jie Hu, Kun Yuan, Minghai Qin, Bahram Afsharmanesh, and Amirali Mirian. Both cataloged 2026 papers are multi-author efforts — first-author leadership on the JEPA training work and a middle position on TierKV — consistent with team-based systems research spanning GPU training infrastructure and mobile inference.
Research keywords
Highlights
First-author JEPA paper claims 1.7x training speedup from execution restructuring alone
Co-authored TierKV tiers mobile LLM KV cache via a predictive solver
Timeline
- 2026Publishes first-author arXiv preprint 'Mask-Aware Execution for Efficient JEPA Training', claiming a 1.7x JEPA training speedup via execution restructuring alone.
- 2026Co-authors 'TierKV: Long-Context On-Device LLMs via Predictive Multi-Tier KV Caching', reporting a 17.6x speedup against MNN-LLM on mobile hardware.
- 2026Both cataloged papers appear in the same year, covering the two halves of the reported field distribution — GPU/HPC training systems and mobile/edge machine-learning 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 in this profile derive from OpenAlex-based public bibliographic data only; no institutional or funding records were used.
- Only two 2026 arXiv preprints are in the reviewed catalog, neither with a confirmed peer-reviewed venue, so the output, impact, and activity dimensions are marked missing.
- Institution and country are absent from the input data; the profile is name-resolved from ORCID 0009-0009-8431-3699 and public bibliographic records.
- The researcher's own name appears within the frequent-coauthor list and two names repeat; collaborator counts were interpreted exactly as given in the raw data.