Hongfan Wu

Hong Kong University of Science and Technology · Hong Kong SAR China

Unclaimed · AI-structured profileOpenAlex

HKUST systems researcher: GPU cluster scheduling & LLM inference serving — fragmentation-aware training, quantized-MoE routing.

papers here

reviewed by Evolve

first-author

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 output score yet: the OpenAlex record shows h-index and i10-index both at 0, with indexed works appearing only in 2026. This axis would measure cumulative publication volume and consistency over time, which cannot be assessed for a two-paper, single-year footprint.

Impact—

No impact score yet: citation counts, h-index, and i10-index are unavailable (both indices recorded at 0), and both papers are recent arXiv preprints. This axis would measure community uptake through citations and derived indicators such as field-weighted influence.

Focus0/100

The focus score of 0 reflects an even split in the raw field distribution: 50% systems / GPU cluster scheduling and 50% machine learning systems (LLM inference serving). OpenAlex topics are similarly dispersed — Cloud Computing and Resource Management (2 works) next to four single-count topics — so no subfield yet dominates this early two-paper record.

Collaboration28/100

The collaboration score of 28 rests on a small raw footprint: 3 distinct coauthors on record — Yukai Zhou, Zhenghong Huang, and Jiheng Zhang. Wu is last author (not corresponding) on COMPASS-ABS and a middle author (not corresponding) on the routing paper, a pattern consistent with deep contribution inside one compact team rather than an independent, multi-group network.

Activity32/100

Activity scores 32, grounded in raw recency counts: 2 works in the last three years, all falling in 2026 per the OpenAlex counts-by-year data (2026: 2). With no earlier indexed output, the publication cadence cannot be judged — one active year, two arXiv preprints, and no multi-year history yet.

Reproducibility55/100

Reproducibility averages 55 across the 2 reviewed papers (reproPapers: 2), both rated medium readiness. COMPASS-ABS ships a proven SIF≤2/N guarantee, but Evolve flags evaluation breadth, migration-overhead accounting, and baseline fairness; the routing paper is rigorously scoped with strong theory and honesty, yet its 2.5% FWP gain on 88 prompts lacks variance estimates — theory-forward, evidence-thin.

Research summary

Hongfan Wu is a researcher at Hong Kong University of Science and Technology whose recent work sits at the intersection of systems and machine learning: GPU cluster scheduling for deep learning training, and serving infrastructure for large language models. The OpenAlex-indexed record is concentrated in 2026, with two arXiv preprints in the Evolve catalog and a topic profile led by Cloud Computing and Resource Management (2 works) alongside single counts in topic modeling, neural network applications, parallel optimization, and distributed computing. The first paper, COMPASS-ABS, proposes a history-free fragmentation metric (SIF) and an anchor/slot/pool scheduler with a proven SIF≤2/N guarantee under a workload-composition condition; Evolve's review calls it conceptually sound and well-evaluated while flagging evaluation breadth, migration overhead accounting, and baseline fairness. The second formulates quality-constrained request routing over a fixed pool of quantized Mixture-of-Experts instances via linear programming duality, reporting a 2.5% FWP gain on 88 prompts without variance estimates. Wu appears as last author (non-corresponding) on COMPASS-ABS and middle author (non-corresponding) on the routing paper, within a small collaborator set of 3 distinct coauthors. Both papers carry medium reproducibility readiness — theory-forward and honestly scoped, but with gaps in statistical reporting and evaluation coverage.

Collaboration pattern

The record shows a small, stable collaboration footprint: 3 distinct coauthors on OpenAlex, with Yukai Zhou, Zhenghong Huang, and Jiheng Zhang as the recurring names. Wu contributes as last author on COMPASS-ABS and a middle author on the MoE routing paper, neither as corresponding author — a pattern consistent with senior-position contributions within one compact systems group rather than an independent, multi-team collaboration network.

Research keywords

GPU cluster schedulingresource fragmentationdeep learning traininggang schedulingbin packingcheckpoint-based migrationMixture-of-ExpertsLLM inference servingpost-training quantizationrequest routinglinear programming dualityquality-degradation budgets

Highlights

History-free fragmentation metric for shared GPU clusters
COMPASS-ABS introduces SIF, a fragmentation metric requiring no historical workload data, paired with an anchor/slot/pool scheduler. Evolve's review judges it conceptually sound and well-evaluated, with a proven SIF≤2/N guarantee under a workload-composition condition, while cautioning that evaluation breadth, migration overhead accounting, and baseline fairness need scrutiny.
Rigorously scoped routing over fixed quantized MoE instances
Frames request routing across a fixed pool of quantized Mixture-of-Experts instances as a quality-constrained optimization solved with linear programming duality. Evolve praises the theory and honesty of the scoping, but notes the reported 2.5% FWP gain on 88 prompts lacks variance estimates.

Timeline

2026COMPASS-ABS preprint appears on arXiv, proposing the SIF fragmentation metric and an anchor/slot/pool scheduler with a proven SIF≤2/N guarantee; Wu is last author.
2026Co-authors a quality-constrained routing scheme over a fixed pool of quantized MoE instances, formulated via linear programming duality; Wu is a middle author.
2026OpenAlex records 2 works in the year, splitting evenly between GPU cluster scheduling and LLM inference serving (50/50 field distribution).
2026Early-stage profile crystallizes: both catalog papers rated medium reproducibility readiness, while citation-based impact signals remain unavailable.

Papers in the Evolve catalog

Frequent co-authors

Yukai Zhou×1 Zhenghong Huang×1 Jiheng Zhang×1

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 derived from OpenAlex only; no ORCID is linked, so this profile is name-resolved from public bibliographic data.
Impact signals are absent: h-index and i10-index read 0 and no citation counts are provided; both papers are 2026 arXiv preprints with no shown peer-review status.
The catalog covers only 2 reviewed papers, so topic, field, and radar readings (collaboration 28, activity 32, focus 0, reproducibility 55) are early-stage and may shift as more output is indexed.
All timeline entries derive from the two provided paper years (both 2026); no biographical or career facts beyond the input data are asserted.