LLM post-training researcher behind GVPO++, spanning policy optimization and self-improving data preparation
2
papers here
1
reviewed by Evolve
0
first-author
0
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—
The output axis is unscored: this snapshot lacks a verifiable publication-count record tied to the ORCID, so total scholarly output cannot be measured honestly. It would aggregate paper counts, venue tiers, and year-by-year output volume.
Impact—
The impact axis is unscored: no citation counts, h-index, or reference-level citation data are available in this snapshot. It would measure citations, citation velocity, and influence on downstream research.
Focus100/100
Focus scores a perfect 100 because the field distribution in the snapshot contains exactly one field — machine learning (LLM post-training / reinforcement learning) — with no scatter across disciplines. The metric rests on a single field-tagged paper, GVPO++, whose keywords (GRPO, on-policy distillation, KL-constrained reward maximization) all sit in that one lane; DataFoundry's data-preparation work is adjacent but not yet field-tagged.
Collaboration51/100
The collaboration axis scores 51 (moderate) from a raw record of 12 distinct co-authors. The roster is one dense cluster — Kaichen Zhang, Yuzhong Hong, Junwei Bao, Hongfei Jiang, Yang Song, Dingqian Hong, Cehao Yang, Xiaojun Wu, Xueyuan Lin, Chengjin Xu, Xuhui Jiang, and Jian Guo — consistent with team-written LLM engineering papers, with Xiong listed last on GVPO++ and mid-list on DataFoundry. The moderate score reflects limited verified co-authorship history rather than an absence of partners.
Activity—
The activity axis is unscored: the snapshot lacks a year-by-year publication record, so research cadence and continuity cannot be assessed. It would track publications per year and the recency of ongoing work.
Reproducibility100/100
Reproducibility scores 100: the mean readiness across Evolve-rated papers is 100, computed from exactly one rated paper. That paper, GVPO++, received a high reproducibility-readiness rating in Evolve's review. The perfect score is real but thin — a single rated paper cannot yet establish a track record, and the same review cautions that the reported 7B-scale gains lack error bars.
Research summary
Hui Xiong (ORCID 0000-0001-6016-6465) is a machine-learning researcher whose cataloged work centers on post-training for large language models, specifically reinforcement-learning-based policy optimization. The Evolve-reviewed 2026 arXiv paper GVPO++, on which Xiong is the last-listed author, extends Group Variance Policy Optimization to on-policy distillation: its zero-sum weighting removes importance sampling while retaining exact-optimum theory under KL-constrained reward maximization. Evolve's review calls the construction elegant and rates reproducibility readiness high, while cautioning that the reported 7B-scale gains lack error bars. A second 2026 arXiv paper, DataFoundry: Evolving Data Preparators via Recursive Self-Improvement, lists Xiong among middle-position co-authors and extends the portfolio toward data preparation and recursive self-improvement, though it carries no Evolve review yet. The bibliometric snapshot behind this profile is narrow: the field distribution tags a single paper under machine learning (LLM post-training / reinforcement learning), producing a perfect focus score of 100, and reproducibility also scores 100 — a mean readiness of 100 computed from exactly one rated paper. Twelve distinct co-authors appear in the record, including Kaichen Zhang, Yuzhong Hong, Chengjin Xu, Xuhui Jiang, and Jian Guo, yielding a moderate collaboration score of 51. Output, impact, and activity axes lack sufficient underlying data and are intentionally left unscored.
Collaboration pattern
Hui Xiong's verified record shows 12 distinct co-authors, giving a moderate collaboration score of 51. The frequent-collaborator list is one dense cluster — Kaichen Zhang, Yuzhong Hong, Junwei Bao, Hongfei Jiang, Yang Song, Dingqian Hong, Cehao Yang, Xiaojun Wu, Xueyuan Lin, Chengjin Xu, Xuhui Jiang, and Jian Guo — consistent with team-written LLM engineering papers. Xiong sits last among authors on GVPO++ and mid-list on DataFoundry.
Research keywords
Highlights
GVPO++ removes importance sampling with exact-optimum theory
DataFoundry pushes toward self-improving data preparation
A second 2026 arXiv catalog entry places Hui Xiong among middle-position co-authors on DataFoundry, which evolves data preparators via recursive self-improvement. No Evolve verdict, keywords, field tag, or reproducibility rating exist for it yet, so this line of work is only partially documented in the catalog.
A perfect but thin reproducibility signal
Timeline
- 2026GVPO++ appears on arXiv with Hui Xiong as last-listed author; Evolve reviews it, rating reproducibility readiness high while noting the 7B-scale gains lack error bars.
- 2026DataFoundry: Evolving Data Preparators via Recursive Self-Improvement enters the catalog with Xiong as a middle-position co-author; no review verdict is available yet.
- 2026Both cataloged papers fall in this single year, making 2026 the only year represented in the current Evolve snapshot for this researcher.
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 are computed from OpenAlex-based public bibliometric data; citation-level and venue coverage may be incomplete.
- "Hui Xiong" is a common name: this profile is name-resolved from public bibliographic data via ORCID 0000-0001-6016-6465, and institution and country are absent from the snapshot.
- Output, impact, and activity axes are intentionally unscored because the underlying counts are missing from this snapshot.
- Only GVPO++ carries a field tag, Evolve verdict, keywords, and reproducibility rating; DataFoundry is cataloged without review metadata, so most conclusions rest on a single reviewed paper.