Systems-minded researcher at UT Austin working on efficient video generation and faster LLM inference.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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 is available yet: the underlying data lacks a publication count and venue list for this researcher. This axis would measure total research output — number of papers, publication venues, and yearly volume.
Impact—
No impact score is available because citation-based signals are missing from the current data. This axis would measure research influence — citations, citing works, and downstream adoption of the researcher's methods.
Focus0/100
The focus score is 0, reflecting the raw field distribution in the reviewed sample: an even 0.5/0.5 split between machine learning for efficient video generation and natural language processing for efficient LLM inference. Within this two-paper sample there is no single dominant topical cluster — the work deliberately spans video and language efficiency rather than concentrating on one subfield.
Collaboration60/100
The collaboration score is 60. The two reviewed papers involve 19 distinct coauthors, with 12 recurring collaborators including Haocheng Xi, Xiuyu Li, and Kurt Keutzer. Authorship position is middle on both papers, neither corresponding — consistent with a contributor role inside larger systems-oriented teams rather than small-team leadership.
Activity—
No activity score is available: the data lacks a full year-by-year publication record, so recent research cadence cannot be computed. This axis would measure publishing frequency and momentum in the most recent years.
Reproducibility55/100
The reproducibility score is 55, the mean over 2 papers with repro-readiness rated medium in both cases. Evolve notes Video DeltaNet decomposes speedups honestly but validates quality only with proxy-VQA, while Osprey reports solid ablations yet single-run evidence — artifacts look reusable, with main results not yet independently confirmed.
Research summary
Chenfeng Xu is a researcher at The University of Texas at Austin whose reviewed work concentrates on making large generative models cheaper to run, along two parallel fronts. The Evolve catalog holds two 2026 arXiv papers, and the underlying field distribution splits evenly — 0.5 machine learning for efficient video generation, 0.5 natural language processing for efficient LLM inference. The first, Video DeltaNet, adds a frame-wise delta-rule linear branch to video diffusion for livestream generation; Evolve credits strong systems work and an honest decomposition of the speedups while noting quality evidence is limited to proxy-VQA. The second, Osprey, warm-starts speculative-decoding drafters from pretrained small language models, reporting +16-23% acceptance gains across three targets with solid ablations, though the evidence is single-run. Both papers place Xu in a middle author position, neither as corresponding author, within larger systems-oriented teams: 19 distinct coauthors appear across the two works, with 12 recurring collaborators including Kurt Keutzer, Haocheng Xi, and Xiuyu Li. Radar scores capture this shape — collaboration 60, reproducibility 55 (mean over both papers, each rated medium for repro readiness), and focus 0, reflecting the even two-field split. Output, impact, and activity axes are not yet scored because the underlying bibliometric signals are missing; no citation counts or venue-level record were available. The profile is name-resolved from public bibliographic data.
Collaboration pattern
Chenfeng Xu's reviewed work is collaborative: 19 distinct coauthors across the two 2026 papers, with a stable core of 12 recurring collaborators — including Haocheng Xi, Xiuyu Li, Kurt Keutzer, Yiming Xie, and Zhaoyang Lv. Both papers list Xu in a middle author position, not corresponding, suggesting contribution within larger systems-oriented teams spanning video generation and LLM inference. The collaboration dimension scores 60.
Research keywords
Highlights
Video DeltaNet: frame-wise delta-rule linear attention for livestream video diffusion
Osprey: target-agnostic pre-training yields stronger speculative drafters
An even two-field split inside one year
The two 2026 reviewed papers divide evenly between efficient video generation and efficient LLM inference (field distribution 0.5/0.5), both with middle authorship and neither corresponding, and both rated medium for reproducibility readiness.
Timeline
- 2026Video DeltaNet appears on arXiv, introducing a video-native hybrid attention with a frame-wise delta-rule linear branch for livestream video generation.
- 2026Osprey appears on arXiv, warm-starting speculative-decoding drafters via target-agnostic pre-training, with reported +16-23% acceptance gains across three targets.
- 2026Both cataloged papers land in 2026, splitting evenly between efficient video generation and efficient LLM inference, with 19 distinct coauthors across the two works.
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.
- Radar dimensions are computed from public bibliometric data only; the underlying publication record may be incomplete.
- Output, impact, and activity axes are unscored because the underlying bibliometric signals are not available.
- No citation counts, h-index, or venue-level publication list were provided; both reviewed works are 2026 arXiv preprints.
- ORCID and country are not provided; the profile is name-resolved from public bibliographic data.