Frank Li
—
LLM inference serving researcher with first-author work on KV-cache sharing and hybrid-state recovery (vLLM, LMCache).
2
papers here
2
reviewed by Evolve
2
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—
No underlying data yet: publication counts, venue record, and yearly totals are absent; this snapshot covers only the two cataloged arXiv papers. This axis would measure total publications and the output trajectory over time.
Impact—
Insufficient data: no citation counts, h-index, or influence indicators exist in the snapshot, so influence on the LLM inference systems community cannot be assessed. This axis would measure citations, citation trajectory, and uptake of the published work.
Focus0/100
The recorded focus score is 0, with a field distribution split evenly (0.5 and 0.5) between 'systems — LLM inference infrastructure' and 'ML systems (LLM inference serving).' In practice both cataloged papers address the same problem area — KV-cache management for vLLM and LMCache — so the two fields are adjacent labels for one engineering focus; the zero score should be read against this narrow two-paper sample.
Collaboration0/100
The collaboration dimension is scored 0, matching a snapshot that records zero distinct co-authors and an empty frequent-coauthor list. Both cataloged papers position Frank Li as first author and neither as corresponding, consistent with solo-authored or single-lead engineering studies. With no institutional affiliation on file, any external collaboration network — labs, advisors, or industry groups — is invisible to this profile, so the zero reflects missing network coverage as much as confirmed solitary work.
Activity—
Not yet measurable: the snapshot spans only 2026 papers, so publication cadence, recent-output trends, and long-run activity levels are unknown. This axis would measure publishing frequency and sustained research activity across recent years.
Reproducibility38/100
Evolve's reproducibility readiness averages 38 across exactly two reviewed papers (reproPapers: 2, reproMean: 38). The shared-KV-caching paper rates medium: its stream-ordering fault diagnosis is rigorous, but evidence spans one host, one model, and two block pairs. The hybrid-state cache-recovery study rates low: bounded but single-model evidence with no code release. Together they suggest careful diagnostic method paired with narrow experimental scope and limited artifact availability.
Research summary
Frank Li's Evolve profile is early-stage, built from two first-author arXiv papers dated 2026. Both target the LLM inference serving stack — specifically KV-cache management around vLLM and LMCache. The first, 'Shared KV Caching for Replicated 27B Inference,' reports a claimed reduction in time-to-first-token from 31.7 to 0.6 seconds via shared KV caching, and diagnoses correctness failures rooted in CUDA stream ordering and pinned host memory; Evolve's review rates the stream-ordering fault diagnosis rigorous while noting the evidence spans one host, one model, and two block pairs, with medium reproducibility readiness. The second, 'Validating Hybrid-State Cache Recovery for GLM-5.3-Flash,' validates a strict-prefix cache-recovery fix for hybrid vLLM+LMCache serving involving hybrid linear attention and checkpoint alignment; the review finds the bounded evidence convincing but limited to one model, with no code release and low reproducibility readiness. Evolve's aggregate reproducibility score for these two papers is 38. The field distribution splits evenly (0.5/0.5) between 'systems — LLM inference infrastructure' and 'ML systems (LLM inference serving).' The bibliometric snapshot records zero distinct co-authors and no frequent collaborators, so the collaboration dimension is scored 0, while output, impact, and activity metrics are missing. Institution, country, and ORCID are not on file, and because 'Frank Li' is a common name, this profile is name-resolved from public bibliographic data.
Collaboration pattern
Within Evolve's snapshot, Frank Li shows a solitary collaboration pattern: zero distinct co-authors, an empty frequent-coauthor list, and collaboration dimension score 0. Both papers are first-author works with corresponding-author status false, consistent with independent or single-lead engineering studies on the vLLM/LMCache stack. Because no institution, country, or ORCID is on file, lab or industry affiliations that might host collaborators cannot be ruled out — the pattern rests on a two-paper sample.
Research keywords
Highlights
Shared KV caching for replicated 27B inference: 31.7s → 0.6s TTFT, with correctness caveats
Validating hybrid-state cache recovery for GLM-5.3-Flash on vLLM+LMCache
Timeline
- 2026First cataloged first-author arXiv paper appears: 'Shared KV Caching for Replicated 27B Inference,' reporting 31.7→0.6s TTFT gains and stream-ordering correctness failures.
- 2026Second first-author arXiv paper follows: 'Validating Hybrid-State Cache Recovery for GLM-5.3-Flash with vLLM and LMCache,' a strict-prefix cache-recovery case study.
- 2026Both papers enter the Evolve catalog with reviews; aggregate reproducibility readiness is scored 38 across the two, rated medium and low respectively.
Papers in the Evolve catalog
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.
- Profile is name-resolved from public bibliographic data; 'Frank Li' is a common name, and institution, country, and ORCID are not on file.
- Evidence covers only the two Evolve-cataloged 2026 arXiv papers; broader publication history, citations, and venue record are not reflected.
- Output, impact, and activity dimensions are reported as missing; the 0 scores for collaboration and focus come from a sparse two-paper snapshot and should be read cautiously.
- Reproducibility score (38) averages Evolve reviews of exactly two papers, one rated low and one medium; single-model and single-host evidence bounds apply to both.