Warwick researcher bridging learning-theoretic optimization with diffusion models and parameter-efficient fine-tuning
2
papers here
2
reviewed by Evolve
0
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—
Output data are insufficient to score this axis. It would measure total publications, annual output and venue tiers; with only two 2026 arXiv preprints reviewed so far, no publication-count or venue-based output score can be computed.
Impact—
Impact cannot be scored: the current snapshot provides no citation counts, h-index or citing works. This axis would measure the citation influence of the reviewed preprints once bibliometric citation data accumulates.
Focus0/100
The field distribution splits exactly 50/50 — machine learning theory (optimization and learning theory) and machine learning (diffusion generative models, parameter-efficient fine-tuning) each weigh 0.5 — producing a focus score of 0: a balanced dual-track portfolio of one theory paper and one applied generative-modelling paper, not single-subfield concentration.
Collaboration41/100
Collaboration scores 41, based on 7 distinct coauthors in the reviewed record; recurring names include Jiayang Gu, Zheng Fang, Lichaun Xiang, Xu Cai, Hongkai Wen, Alex Buna and Patrick Rebeschini. Both 2026 papers place Liu in a middle author position without corresponding role, consistent with contributor roles inside small theory and generative-modelling teams.
Activity—
Activity data are insufficient to score this axis. It would measure publication cadence and year-over-year continuity; both reviewed items are dated 2026, so no multi-year distribution exists from which to compute an activity score.
Reproducibility78/100
Reproducibility is the strongest recorded dimension at 78, the mean over 2 reviewed papers. The ridge gradient descent compute-allocation paper carries high readiness, reflecting explicit spectral-rate analysis in linear full-batch settings, while PRISM-LoRA is rated medium — real but uneven gains with thin statistics — so the score rests on one strong and one moderate anchor.
Research summary
Fanghui Liu is a machine learning researcher at the University of Warwick whose reviewed work splits evenly across two tracks: machine learning theory (optimization and learning theory) and machine learning (diffusion generative models, parameter-efficient fine-tuning), each weighted 0.5 in the field distribution. The Evolve catalog holds two 2026 arXiv preprints, both with Liu in a middle author position. 'Compute-Optimal Pretrain–Fine-tune in Ridge Gradient Descent' derives the optimal pretrain–fine-tune compute split in ridge gradient descent through bottleneck spectral rates; Evolve assesses the theory as rigorous but confined to linear full-batch GD settings, with high reproducibility readiness. PRISM-LoRA ('Principal-timestep Restricted Init via Sparse Matrix-decomposition in Flow-matching') proposes timestep-restricted sparse initialization to counter spectral-LoRA failure in flow matching; the gains are real but uneven and the statistics thin, yielding medium readiness. Together the two papers give a mean reproducibility score of 78. Collaboration scores 41 over 7 distinct coauthors, with Jiayang Gu, Zheng Fang, Lichaun Xiang, Xu Cai, Hongkai Wen, Alex Buna and Patrick Rebeschini recurring. Output, impact and activity metrics are missing from the underlying bibliometric data, so this profile is an early-stage, name-resolved snapshot of an active 2026 research period rather than a full career record.
Collaboration pattern
Fanghui Liu's reviewed record shows 7 distinct coauthors and a collaboration score of 41. Recurring collaborators include Jiayang Gu, Zheng Fang, Lichaun Xiang, Xu Cai, Hongkai Wen, Alex Buna and Patrick Rebeschini. In both 2026 papers Liu sits in a middle author position without corresponding-author status, suggesting a contributor-style pattern within small teams spanning learning theory and diffusion-based generative modelling rather than a solo or corresponding-lead profile.
Research keywords
Highlights
Optimal pretrain–fine-tune compute split proven for ridge gradient descent
PRISM-LoRA: timestep-restricted sparse initialization for flow matching
Evenly balanced dual-track record across theory and generative modelling
The reviewed field distribution weights machine learning theory (optimization, learning theory) and machine learning (diffusion generative models, parameter-efficient fine-tuning) at 0.5 each — one rigorous optimization-theory paper alongside one diffusion/LoRA paper, both 2026 arXiv preprints with middle-author placement.
Timeline
- 2026Posts 'Compute-Optimal Pretrain–Fine-tune in Ridge Gradient Descent' to arXiv, deriving the optimal pretrain–fine-tune compute split in ridge GD via bottleneck spectral rates.
- 2026Introduces PRISM-LoRA on arXiv, using principal-timestep restricted sparse matrix-decomposition initialization to address spectral-LoRA failure in flow matching.
- 2026Both preprints enter the Evolve catalog with reproducibility readiness ratings of high and medium respectively, and Liu credited as middle author on each.
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 dimensions are computed from OpenAlex-derived data only; the output, impact and activity axes lack sufficient data and are unscored.
- Coverage is limited to two 2026 arXiv preprints reviewed by Evolve; peer-reviewed venue history is not represented.
- Both reviewed papers list Fanghui Liu as a middle author with no corresponding-author role.
- Profile is name-resolved from public bibliographic data linked to ORCID 0000-0003-4133-7921; the institution's country is not recorded.