First-author systems researcher at UC Berkeley working on elastic LLM inference serving and GPU memory management
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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 output data is available yet: the catalog holds only the two reviewed 2026 papers, with no publication counts, venue history, or yearly output series. This axis would measure cumulative and per-year research volume.
Impact—
No citation counts, h-index, or downstream-adoption signals were provided, so research impact cannot be scored. This axis would measure citations, reuse of the reported 16-17% and 52% gains by later systems, and uptake in the LLM-serving community.
Focus0/100
The focus score is recorded as 0 and should be read cautiously against its raw basis: the field distribution splits evenly (0.5/0.5) between GPU memory management for LLM serving and LLM inference serving, two tightly adjacent subfields. With only two reviewed papers in the catalog, this axis is weakly determined; the raw split suggests a deliberately narrow agenda around elastic LLM-serving systems rather than demonstrated breadth or scatter.
Collaboration50/100
The collaboration score is 50, a mid-range value built on 11 distinct coauthors across the reviewed work. The frequent-coauthor list mixes Berkeley systems leadership (Ion Stoica) with a large industry cluster (Tian Xia, Krishna Malladi, Chunqiang Tang, Adnan Aziz, and others), and the Crossflow review attributes that system to Meta — consistent with industry-partnered systems projects. Both papers are first-authored without corresponding-author flags, positioning Xu as the hands-on lead within sizeable teams.
Activity—
No activity signal is available: the data provides no publication-frequency series, submission cadence, or multi-year timestamps beyond the two 2026 papers. This axis would measure recent publishing rhythm and sustained engagement with the LLM-serving field over time.
Reproducibility38/100
The reproducibility score of 38 is the mean across 2 reproducibility-assessed papers. Crossflow is rated low readiness: its evaluation is well-ablated and honest (16-17% geomean gains), but traces are private and no code is available to verify claims. EMA rates medium: up to 52% throughput gains with a sound design, yet no variance data and no modern-hardware results. The sub-40 mean reflects this gap between careful measurement and externally checkable artifacts.
Research summary
Yi Xu is a computer-systems researcher at UC Berkeley whose reviewed work concentrates on the systems layer of large language model inference. Evolve's catalog holds two 2026 arXiv papers, both with Xu in the first-author position: Crossflow, which lends decode-node slack to prefill in agentic LLM serving and reports 16-17% geomean gains, and EMA, a performance-transparent GPU-to-GPU memory-sharing design reporting up to 52% throughput gains. The raw field distribution splits evenly (0.5/0.5) between GPU memory management for LLM serving and LLM inference serving, indicating a narrow, deliberate agenda rather than scattered interests. Collaboration data records 11 distinct coauthors (score 50), with a frequent-coauthor set that spans Berkeley systems leadership and a large industry cluster; the Crossflow review attributes that system to Meta, pointing to industry-partnered work. The reproducibility score of 38 (mean over 2 papers) reflects honest, well-ablated evaluations held back by private traces, unreleased code, and missing variance data. Output, impact, and activity axes lack sufficient data and are left unscored. This profile is name-resolved from public bibliographic data.
Collaboration pattern
Yi Xu's reviewed work involves 11 distinct coauthors (collaboration score 50, mid-range). The frequent-coauthor set spans Berkeley systems leadership — Ion Stoica — and a dense industry cluster including Tian Xia, Ehsan K. Ardestani, Wenyin Fu, Martin Schatz, Krishna Malladi, Zhan Shu, Adnan Aziz, Shobhit Kanaujia, Ajit Mathews, and Chunqiang Tang; the Crossflow review attributes that system to Meta. Both papers are first-authored without corresponding-author credit, consistent with a hands-on lead role inside large industry-academia teams.
Research keywords
Highlights
Crossflow: lending decode-node slack to prefill in agentic LLM serving
EMA: performance-transparent memory sharing across GPUs
Timeline
- 2026First-authored Crossflow appears on arXiv, introducing prefill-decode elasticity for agentic LLM serving with 16-17% geomean gains.
- 2026First-authored EMA appears on arXiv, reporting performance-transparent cross-GPU memory sharing with up to 52% throughput gains.
- 2026Both reviewed papers land in 2026 with Xu in the first-author position, marking an early-stage but concentrated entry into LLM-serving systems within an 11-coauthor network.
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 scores rest on a very small base: only 2 reviewed papers, both 2026 arXiv preprints; the output, impact, and activity axes are unscored for lack of data.
- 'Yi Xu' is a common name; this profile is name-resolved from public bibliographic data and the reviewed catalog, with no verified ORCID on record.
- Performance figures (16-17% geomean, up to 52% throughput) are quoted verbatim from Evolve's review verdicts and are not independently verified.
- The reproducibility score of 38 is a mean over the 2 assessed papers; readiness ratings reflect code, trace, and hardware-coverage availability at review time.