Researcher across efficient LLM architecture and physics-aware video world models; arXiv work on KV cache compression.
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—
No output data is available yet. This axis measures total publications, annual volume, venue tiers, and authorship positions; the current snapshot contains only two 2026 arXiv papers, so Evolve cannot score it and we do not estimate.
Impact—
No impact data is available yet. This axis measures citations, h-index, and downstream uptake of the researcher's work; with only two very recent 2026 arXiv papers and no citation snapshot, it is left unscored.
Focus0/100
Evolve's focus score is 0/100: across the two reviewed papers, the field distribution splits exactly 50/50 between efficient large language model architecture and computer vision (video generation and embodied world models). On current evidence the portfolio is broad — spanning KV cache compression and physics-aware video generation — rather than concentrated in a single specialty.
Collaboration85/100
This is the strongest measured axis at 85/100: the raw record lists 71 distinct coauthors across only two reviewed 2026 papers. Both are middle-author contributions inside very large teams, and the KV cache paper carries DeepSeek-AI collective authorship — a signature of high-throughput, industrial-scale research where credit is broadly shared rather than individual-led.
Activity—
No activity data is available yet. This axis measures publication cadence and research continuity over multiple years; the record here covers a single year (2026) across two papers, so no trend or rhythm can be scored.
Reproducibility20/100
Reproducibility averages 20/100 over the two scored papers, both rated low readiness. Evolve's verdicts note that DeepSeek-V4.1-Flash is self-evaluated with no ablations and mixed Table 1 results, while StrucPhysVideo relies on single-run evaluations and a private dataset. Releasing code, data, and ablations would be the clearest path to stronger replication confidence.
Research summary
This profile is name-resolved from public bibliographic data: the source record's display name is unresolved (recorded as ":"), and all facts below come from two 2026 arXiv papers reviewed by Evolve plus the co-occurring coauthor list. The researcher works across two fields in near-equal measure — efficient large language model architecture and computer vision focused on video generation and embodied world models — each accounting for 50% of the reviewed output, which yields a focus score of 0/100 on Evolve's scale. On the DeepSeek-V4.1-Flash paper (middle author), the team claims a 552B-parameter mixture-of-experts model matching frontier agents at 1/4–1/8 the KV cache; Evolve calls it engineering-strong but self-evaluated, with no ablations and mixed Table 1 results. On StrucPhysVideo (also middle author), the team reports 45.5% Physics-IQ state of the art via structured physics captions and action-conditioned video models, though Evolve flags single-run evaluations and a private dataset. Both papers carry low reproducibility readiness, averaging 20/100 over the two scored papers. Collaboration is the strongest measured signal: 85/100, with 71 distinct coauthors across just two papers, including DeepSeek-AI collective authorship — a signature of large-team, industrial-scale research. Output, impact, and activity axes lack underlying data and are deliberately unscored.
Collaboration pattern
Both reviewed papers are middle-author contributions to very large teams: the DeepSeek-V4.1-Flash paper carries DeepSeek-AI collective authorship, and the frequent-coauthor list is headed by that collective alongside individuals such as Anyi Xu and B. Li. With 71 distinct coauthors behind just two 2026 papers and a collaboration score of 85/100, the pattern suggests high-throughput industrial or lab-scale research where credit is broadly shared; many frequent coauthors appear with truncated initials, so disambiguation may be noisy.
Research keywords
Highlights
Middle-author role on DeepSeek-V4.1-Flash KV cache compression
Contributor to StrucPhysVideo physics-aware world models
Broad large-team collaboration footprint
Across only two reviewed 2026 papers the record lists 71 distinct coauthors, underpinning Evolve's collaboration score of 85/100; DeepSeek-AI collective authorship on one paper points to industrial-scale team science, though many coauthor names carry truncated initials.
Timeline
- 2026Co-authored (middle author position, not corresponding) the DeepSeek-V4.1-Flash arXiv paper on KV cache compression for a 552B mixture-of-experts model.
- 2026Co-authored (middle author position, not corresponding) the StrucPhysVideo arXiv paper, reporting 45.5% Physics-IQ performance via structured physics captions and action-conditioned video models.
- 2026Bibliometric footprint across the two reviewed papers: 71 distinct coauthors and an even 50/50 field split between efficient LLM architecture and video/embodied world model research.
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.
- The researcher's display name resolved to ":" in the source record; this profile is built purely from public bibliographic data and two Evolve-reviewed papers, without identity verification.
- No institution, country, or ORCID is present in the input data.
- Bibliometric scores derive only from the two 2026 arXiv papers reviewed by Evolve; the output, impact, and activity axes have no underlying data and are unscored.
- Several frequent coauthors are recorded with truncated initials (e.g., B. Li, C. C. Yu), so the 71-coauthor count may include unresolved duplicates.