Ruiyi Fang

University of Electronic Science and Technology of China

Unclaimed · AI-structured profileOpenAlex

Early-stage machine learning researcher at UESTC: graph representation learning and efficient LLM inference

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-13) 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 underlying data on this researcher's total publication volume, paper counts by venue, and authorship positions is not available in the current snapshot. This axis would measure overall research productivity across the cataloged record.

Impact—

Not yet scored: citation counts, h-index-style indicators, and measures of downstream uptake are absent from the current snapshot. This axis would measure the influence of the work on subsequent research; no estimate is made until such data exists.

Focus0/100

Scores 0/100, the lowest concentration band. The raw field distribution explains it: the two cataloged papers split evenly, 0.5 and 0.5, between machine learning for graph representation learning and machine learning for efficient LLM inference, leaving no dominant subfield. An even split across two distinct method communities reads as exploratory breadth typical of an early-stage researcher rather than a settled agenda, and the score should be revisited as the catalog grows.

Collaboration54/100

Scores 54/100, a mid-range value. The raw record shows 14 distinct coauthors across the two cataloged papers, and twelve recurring collaborators are listed, including Xiangyu Wang, Shuo Wang, Zhao Kang, Qiuhao Zeng, Jerry Huang, and Boyu Wang. Both papers place Fang in a middle author position without the corresponding-author role, a pattern consistent with substantive contribution inside larger team-based machine learning groups rather than leading projects as first or corresponding author.

Activity—

Not yet scored: the snapshot lacks per-year publication counts and timing data needed to gauge research tempo. This axis would measure how actively the researcher publishes over time; with only two 2026 papers visible, no publication rate can be estimated.

Reproducibility78/100

Scores 78/100, the mean reproducibility-readiness rating across 2 reviewed papers. MINT, the optimal-transport graph prompting paper, is rated high readiness, while ARM, the routed-memory KV cache paper, is rated medium, with Evolve noting baseline confounds that weaken attribution of its efficiency gains. A 78/100 average on an arXiv-only record signals comparatively thorough empirical reporting, though the evidence base of two papers is too small to treat the score as stable.

Research summary

Ruiyi Fang is a machine learning researcher affiliated with the University of Electronic Science and Technology of China whose reviewed output to date splits evenly between two subfields: machine learning for graph representation learning and machine learning for efficient LLM inference (0.5 each in the cataloged field distribution). This is an early-stage profile anchored by two 2026 arXiv papers reviewed by Evolve. ARM introduces a differentiable routed-memory KV cache with gated writes and adaptive retrieval for long-context inference, using Gumbel-Softmax routing to make sparse attention learnable; Evolve judges the efficiency gains plausibly real but notes baseline confounds that weaken attribution, rating reproducibility readiness medium. MINT formulates graph prompting as a measure-constrained optimal transport problem with fixed marginals, solved via Sinkhorn, targeting few-shot node classification after GNN pretraining, including heterophilic settings; its theory is exact and evaluation broad, although the paper's own controls show near-zero accuracy effect, and readiness is rated high. On both papers Fang appears in a middle author position, not as corresponding author. Collaboration metrics show 14 distinct coauthors and a mid-range collaboration score (54/100), with a recurring circle that includes Xiangyu Wang, Shuo Wang, Zhao Kang, Qiuhao Zeng, Jerry Huang, and Boyu Wang. Output, impact, and activity remain unscored because the underlying bibliometric data is insufficient.

Collaboration pattern

The record shows 14 distinct coauthors across two 2026 arXiv papers, with a mid-range collaboration score of 54/100. Twelve recurring collaborators are listed, including Xiangyu Wang, Shuo Wang, Zhao Kang, Qiuhao Zeng, Jerry Huang, Peng Lu, Gezheng Xu, Charles Ling, Boyu Wang, and Gang Niu. On both papers Fang sits in a middle author position without the corresponding-author role, suggesting contribution within larger team-based machine learning groups rather than solo or lead-author work.

Research keywords

KV cache managementlong-context LLM inferencesparse attentionGumbel-Softmax routingneural memoryadaptive retrievalgraph prompt learningoptimal transportSinkhornfew-shot node classificationGNN pretrainingheterophily

Highlights

MINT: graph prompting as measure-constrained optimal transport
Casts graph prompt learning as a fixed-marginal optimal transport problem solved with Sinkhorn to balance prompt usage. The theory is exact and the evaluation broad, covering few-shot node classification after GNN pretraining including heterophilic graphs, though the paper's own controls show a near-zero accuracy effect.
ARM: learnable sparse attention via routed memory
A differentiable routed-memory KV cache with gated writes and adaptive retrieval for long-context LLM inference, trained with Gumbel-Softmax routing so sparse control becomes learnable. Evolve assesses the efficiency gains as plausibly real while cautioning that baseline confounds weaken attribution.
A methodological signature: making discrete choices differentiable
Across both papers the common thread is turning discrete structural decisions — which KV entries to keep, how prompt mass is allocated — into learnable optimization problems (Gumbel-Softmax routing and Sinkhorn-based optimal transport), with reproducibility readiness rated medium to high.

Timeline

2026Posts ARM on arXiv: a differentiable routed-memory KV cache with gated writes and adaptive retrieval for learnable sparse control in long-context LLM inference (middle author, non-corresponding).
2026Posts MINT on arXiv: casts graph prompting as measure-constrained optimal transport with fixed marginals for few-shot node classification after GNN pretraining (middle author, non-corresponding).
2026Evolve reviews both papers: reproducibility readiness is rated high for MINT and medium for ARM, averaging 78/100 across the 2-paper record.

Papers in the Evolve catalog

Frequent co-authors

Xiangyu Wang×1 Shuo Wang×1 Zhao Kang×1 Qiuhao Zeng×1 Jerry Huang×1 Peng Lu×1 Gezheng Xu×1 Zihao Jing×1 Yufei Cui×1 Charles Ling×1 Gang Niu×1 Boyu Wang×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.