UCL researcher applying information-theoretic objectives to continual learning and RL for large language models
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) 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: this snapshot lacks publication-volume data such as paper counts and yearly output. This axis would measure how much and how steadily this researcher publishes over time.
Impact—
Not yet scored: citation-based measures (citation counts, h-index, influence) are absent from the available data. This axis would quantify how widely the work is taken up by other researchers.
Focus0/100
The focus score is 0, reflecting a perfectly even split in the field distribution: 50% machine learning (continual learning for LLMs) and 50% machine learning — reinforcement learning for large language models. This is breadth, not drift: both halves pursue the same information-proximal idea — I-projection-bounded self-distillation in iSDFT and information-density re-timing of PPO in InfoPPO — so the low score captures symmetric two-track output rather than scattered topics.
Collaboration59/100
Collaboration scores 59/100, with 18 distinct coauthors across the reviewed record and twelve recurring collaborators. One cluster spans Ahmed Khaled Khamis, Xiaotong Ji, Hassan Jaber, Rasul Tutunov, Matthieu Zimmer, and Haitham Bou-Ammar; a second spans Yongcheng Zeng, Xinyu Cui, Yan Song, Guoqing Liu, Hongsheng Xin, and Kaike Zhang. Authorship sits mid-list on iSDFT and last on InfoPPO, neither corresponding — the pattern of a contributing team member across two collaborating groups rather than a solo project lead.
Activity—
Not yet scored: recent-activity signals (publication recency and cadence) are unavailable beyond the two 2026 arXiv preprints. This axis would measure ongoing research tempo in the most recent years.
Reproducibility78/100
Reproducibility averages 78/100 across the two Evolve-reviewed papers (reproMean 78 over reproPapers 2). The iSDFT self-distillation paper is rated high readiness and InfoPPO medium. Both verdicts flag single-seed evaluation evidence — iSDFT additionally caps its math evaluations — so the score reflects well-specified, checkable methodology whose gains rest on limited statistical evidence rather than confirmed multi-seed replication.
Research summary
Jun Wang is a machine learning researcher affiliated with University College London whose reviewed work concentrates on information-theoretic training methods for large language models. The Evolve catalog contains two 2026 arXiv preprints. The first, iSDFT: Information-Proximal Self-Distillation for Continual Learning in LLMs, frames continual learning as a budgeted information problem, using KL I-projection to cap teacher information during on-policy self-distillation; Evolve notes improvements across four LLM backbones while flagging single-seed evidence and capped math evaluations. The second, Information-Time Proximal Policy Optimization (InfoPPO), re-times PPO updates by information density, with reported gains on Qwen3 mathematical tasks; Evolve credits its coherent theory while tempering claims over the same single-seed limitation. The field distribution splits evenly — 50% machine learning (continual learning for LLMs) and 50% machine learning — reinforcement learning for large language models — yielding a focus score of 0, although both halves share an information-proximal design philosophy. Collaboration scores 59/100 across 18 distinct coauthors with twelve recurring collaborators, and reproducibility averages 78/100 over the two papers, rated high and medium readiness respectively. Output, impact, and activity metrics are not yet available in this snapshot. Because 'Jun Wang' is a common name, this profile is name-resolved from public bibliographic data.
Collaboration pattern
Wang's reviewed record links to 18 distinct coauthors (collaboration score 59/100), with twelve recurring collaborators. One cluster spans Ahmed Khaled Khamis, Xiaotong Ji, Hassan Jaber, Rasul Tutunov, Matthieu Zimmer, and Haitham Bou-Ammar; a second spans Yongcheng Zeng, Xinyu Cui, Yan Song, Guoqing Liu, Hongsheng Xin, and Kaike Zhang. Holding a middle position on iSDFT and the final position on InfoPPO — neither corresponding — suggests a contributing team member across two collaborating groups rather than a solo project lead.
Research keywords
Highlights
Budgeted teacher information for continual LLM learning
Re-timing PPO by information density
A strong reproducibility signal across a young record
Across the two reviewed preprints, mean reproducibility is 78/100 — high readiness for iSDFT and medium for InfoPPO — with single-seed evidence cited as the main tempering factor in both verdicts.
Timeline
- 2026iSDFT: Information-Proximal Self-Distillation for Continual Learning in LLMs posted to arXiv, with Wang in a middle authorship position.
- 2026Information-Time Proximal Policy Optimization posted to arXiv, with Wang in the last authorship position.
- 2026Both preprints enter the Evolve catalog, with reproducibility readiness rated high and medium respectively.
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 derive from an OpenAlex-based public snapshot only; citation counts, h-index, and publication-volume data are absent, so output, impact, and activity axes are unscored.
- 'Jun Wang' is a common name; this profile is name-resolved from public bibliographic data and may reflect aggregation or misattribution across records.
- Qualitative assessments rest on only two Evolve-reviewed 2026 arXiv preprints; the reproducibility score (78) is computed over just those two papers.
- No ORCID or country is provided in the source data; the affiliation (University College London) is as listed and unverified.