Feng Liu

—

Unclaimed · AI-structured profileOpenAlex

Researcher spanning audio representation learning and scientific ML, with controlled empirical studies of pre-training

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) 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: no publication-count data underlies this axis in the current snapshot. It would measure total scholarly output — number of papers over time and by venue — which a record limited to three reviewed 2026 arXiv papers cannot establish.

Impact—

Not yet scored: citation-based measures such as citation counts, h-index, or citing works are absent from the snapshot. This axis would quantify how widely the work is used by other researchers; no such data is available yet.

Focus0/100

Focus scores 0 on the radar, driven by the underlying field distribution: audio representation learning and scientific machine learning / CFD each account for exactly 50%. The three reviewed papers span these two distant domains plus a recommender-systems contribution, so output is deliberately broad rather than concentrated. The low score signals breadth across subfields, not low quality, and would rise if future work consolidated around one line.

Collaboration58/100

Collaboration scores 58, grounded in a network of 17 distinct co-authors of whom 12 recur as frequent collaborators, including Chengqi Yang, Yiran Qiao, Xingyu Lou, Xiang Ao, and Jun Wang. Authorship position is consistent across the three reviewed papers — last author on two, middle author on one, never corresponding — fitting a senior-in-team contribution style that spans an audio group, a scientific-ML group, and an industry collaboration.

Activity—

Not yet scored: the snapshot lacks full publication-date coverage, so recent productivity cadence cannot be computed. This axis would measure output rate over recent years; at present only three 2026 papers are on record, which is not enough to establish a trend.

Reproducibility78/100

Reproducibility scores 78, the mean across the 2 papers Evolve could score, and the pattern is uneven. The PINN automatic-differentiation-versus-discretization study earns high readiness with solid, well-controlled evidence; the procedural-audio pre-training study is medium, with thin statistics; the mobile-game user-representation paper is low, with data and code closed. The average reflects a generally transparent empirical style held back by one closed-source industry collaboration.

Research summary

Feng Liu is a machine learning researcher whose reviewed output splits evenly between two areas — audio representation learning and scientific machine learning / computational fluid dynamics — with additional work on user behavior modeling for recommender systems. Evolve's catalog covers three 2026 arXiv papers. In two Liu is the last author, and in none the corresponding author. 'Rethinking Procedural Audio Pre-training: Source Scaling and Objective Adaptation' is a controlled study finding that procedural-audio scale is multidimensional and that masking must adapt to the source, with plausible but modest gains and thin statistics. 'A Systematic Analysis of Automatic Differentiation versus Discretization-based Constraints for Physics-Informed PDE Solvers' is judged to have solid, well-controlled evidence, though regime boundaries remain qualitative and 2D-steady only. As a middle author, Liu contributed to 'User Representation via Cross Multi-source Behavior Pre-training for Mobile Games', which reports wins on 5 of 6 game tasks (+1.05–2.85 AUC points) with data and code closed. Radar scores mirror this pattern: reproducibility 78 (mean of 2 scored papers), collaboration 58 (17 distinct co-authors), and focus 0, consistent with the 50/50 field split. Output, impact, and activity are not yet scored.

Collaboration pattern

The network contains 17 distinct co-authors, with 12 recurring frequently — including Chengqi Yang, Yiran Qiao, Xingyu Lou, Xiang Ao, and Jun Wang — giving a collaboration score of 58. Positions across the three reviewed papers are consistent: last author twice, middle author once, and never corresponding author, fitting a senior-in-team contribution style that spans an audio group, a scientific-ML group, and an industry collaboration. Cross-domain co-authorship matches the even 50/50 field split.

Research keywords

Procedural audioSynthetic dataAudio pre-trainingMasked autoencoderPhysics-informed neural networksAutomatic differentiationDiscretization-based constraintsScientific machine learningComputational fluid dynamicsUser representation pre-trainingGraph neural networksGame recommendation

Highlights

Systematic comparison of AD vs. discretization constraints for PINNs
Last-authored empirical analysis of automatic differentiation versus discretization-based PDE constraints (finite difference, finite volume, graph neural networks) in physics-informed solvers. Evolve finds the evidence solid and well-controlled, while noting that regime boundaries remain qualitative and cover 2D-steady cases only; reproducibility readiness is rated high.
Procedural audio pre-training: scale is multidimensional
Last-authored controlled study of procedural audio pre-training (AudioMAE-style masked autoencoders, FDSL synthetic sources) showing that source scaling is multidimensional and masking objectives must adapt to the source. Gains are plausible but modest and the statistics are thin; reproducibility readiness is medium.
Cross-source behavior pre-training for mobile-game user representation
Middle-authored industry work with OPPO and CAS that pre-trains user embeddings from cross-source device logs via a cascaded mask-then-predict scheme (CM-PTM), with industry A/B testing. It wins on 5 of 6 game tasks (+1.05–2.85 AUC points), but data and code are closed and the statistics are thin; reproducibility readiness is low.

Timeline

2026Publishes 'Rethinking Procedural Audio Pre-training: Source Scaling and Objective Adaptation' on arXiv as last author — a controlled study of synthetic-data scaling for audio representation learning.
2026Publishes 'A Systematic Analysis of Automatic Differentiation versus Discretization-based Constraints for Physics-Informed PDE Solvers' on arXiv as last author; Evolve rates its reproducibility readiness high.
2026Contributes as a middle author to 'User Representation via Cross Multi-source Behavior Pre-training for Mobile Games' on arXiv, reporting wins on 5 of 6 game tasks with +1.05–2.85 AUC points.

Papers in the Evolve catalog

Frequent co-authors

Chengqi Yang×1 Yiran Qiao×1 Xingyu Lou×1 Zijun Zhou×1 Xiaoyun Mo×1 Changwang Zhang×1 Jiayuan Xu×1 Jun Wang×1 Xiang Ao×1 Jiajun Peng×1 Fengrui Liu×1 Xinyu Liu×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.

'Feng Liu' is a common name; this profile is name-resolved from public bibliographic data, and no ORCID or affiliation was available to confirm identity.
Radar scores and the field distribution come from a single bibliometric snapshot; no citation counts, h-index, venue list, or full publication history were provided.
Coverage is limited to the three 2026 arXiv papers reviewed by Evolve; other works by this author may exist but were not assessed.
Output, impact, and activity axes lack underlying data (dimMissing), so those scores are deliberately omitted rather than estimated.