Kuaishou researcher bridging statistical inference for LLM evaluation and efficient generative-recommendation serving
2
papers here
2
reviewed by Evolve
1
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 volume is not yet scored: the snapshot provides no aggregate publication or preprint counts for this researcher, so total output cannot be computed. This axis would measure cumulative papers, arXiv preprints, and venue distribution over time.
Impact—
No citation counts or citation-based metrics are available in this snapshot, so impact is not scored. This axis would measure citations to the two reviewed papers and their uptake in the LLM-evaluation and recommendation-serving communities.
Focus0/100
Focus scores 0 on the 0-100 radar, matching the raw field distribution: an exact 0.5/0.5 split between 'statistics for machine-learning evaluation (LLM reasoning benchmarks)' and 'machine learning systems (inference serving)'. With only two Evolve-reviewed papers, one per field, the portfolio reads as two parallel agendas — rigorous pass@k inference on one side, generative-recommendation decode efficiency on the other — rather than a single concentrated line.
Collaboration54/100
Collaboration scores 54 out of 100, computed from 14 distinct coauthors in the snapshot. The frequent-coauthor list carries 12 entries, including Xianyang Zhang, Jun Chen, Xiangrui Yang, Cheng Peng, and Yunfeng Zhao, though it also contains the researcher's own name — a name-resolution artifact. The two reviewed papers show different authorship modes: first author on the statistics paper, middle author on the Kuaishou systems paper.
Activity—
Activity is not scored yet: beyond the two 2026 arXiv papers, the snapshot contains no multi-year publication timeline, so yearly output rate cannot be computed. This axis would measure publication cadence and consistency across recent years.
Reproducibility60/100
Reproducibility scores 60, the mean across the 2 Evolve-assessed papers. The split is stark: the first-authored pass@k inference paper is rated high readiness, described as a pre-registered formal paired-inference framework, while OneLA is rated low, with Evolve flagging its evaluation as sound but confined to a single model. The 60 mean reflects averaging one high and one low rating.
Research summary
Chen Yang is a researcher at Kuaishou (China) whose Evolve-reviewed output spans two complementary areas: statistical inference for machine-learning evaluation and machine-learning systems for inference serving. The bibliometric snapshot splits these evenly, 0.5 versus 0.5, between 'statistics for machine-learning evaluation (LLM reasoning benchmarks)' and 'machine learning systems (inference serving)'. The first-authored 2026 arXiv paper 'RLVR is a Kernel, Not a Function: Statistical Inference for pass@k Crossovers' builds a pre-registered paired-inference framework for pass@k comparisons on LLM reasoning benchmarks, combining simultaneous confidence bands, a multiplier bootstrap, crossover inference, and a Markov-kernel view of RLVR training. Evolve calls the framework rigorous but notes that only one small, token-limit-dependent reversal is statistically confirmed, and rates reproducibility readiness high. The second paper, 'OneLA: Scaling Linear-Attention Decoding to Large Beams in Generative Recommendation' (2026, arXiv, middle authorship), replaces per-beam Gated DeltaNet states with one shared state plus compact transition records, claiming 1.54–2.46× decode speedups for generative recommendation; Evolve judges it sound but based on a single-model evaluation, with low reproducibility readiness. Around this work, the snapshot records 14 distinct coauthors and a mean reproducibility score of 60 across 2 assessed papers. 'Chen Yang' is a common name, so this profile is name-resolved from public bibliographic data rather than a verified identity claim.
Collaboration pattern
The snapshot records 14 distinct coauthors, giving a collaboration score of 54 out of 100. The frequent-coauthor list carries 12 entries — Xianyang Zhang, Jun Chen, Xiangrui Yang, Cheng Peng, Yunfeng Zhao, Liang Zeng, Ao Hu, Jiawei Yang, Shengzhe Wang, Jingshan Lv, and Xiao Liang — with the researcher's own name also appearing, a typical name-resolution artifact. Authorship modes differ across the two papers: first author on the statistics paper, middle author on the Kuaishou inference-serving paper.
Research keywords
Highlights
Statistical inference for pass@k crossovers in RLVR evaluation
Shared-state decoding for large-beam generative recommendation
Timeline
- 2026First-authored 'RLVR is a Kernel, Not a Function: Statistical Inference for pass@k Crossovers' appears on arXiv, introducing a pre-registered paired-inference framework for pass@k crossovers in LLM reasoning benchmarks.
- 2026Co-authors 'OneLA: Scaling Linear-Attention Decoding to Large Beams in Generative Recommendation' on arXiv, listed in a middle authorship position.
- 2026Evolve reviews both papers: the pass@k framework earns high reproducibility readiness while OneLA is rated low on its single-model evaluation, giving a reproducibility mean of 60 across the 2 assessed papers.
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 scores come from public bibliographic data (OpenAlex-style) only; no citation counts, h-index, or venue-level publication history are included.
- 'Chen Yang' is a common name; this profile is name-resolved from public bibliographic data, not a verified identity claim. The researcher's own name also appears within the frequent-coauthor list, a typical bibliographic artifact.
- Qualitative verdicts rest on only 2 Evolve-reviewed arXiv papers, both from 2026; the timeline cannot extend earlier than 2026 from this data.
- The output, impact, and activity dimensions are unscored because the snapshot lacks publication counts, citation data, and multi-year activity records.