Machine-learning researcher spanning efficient LLM inference, LLM agents, and NLP for developer Q&A mining
3
papers here
3
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—
Insufficient data: this snapshot contains no aggregate output metrics such as total publication count, venue distribution, or productivity over time. The axis would measure research output volume and its trajectory once a fuller bibliographic record is linked to this profile.
Impact—
No citation data is available for this profile, so impact cannot be scored. The axis would measure citation-based influence — accumulated citations, aggregate indices, and uptake of the three reviewed 2026 preprints once citation sources are connected.
Focus0/100
The focus axis scores 0, reflecting an evenly split field distribution: efficient LLM inference, LLM agents, and NLP for Q&A/software-repository mining each account for roughly 33% of the reviewed papers. The strands are related — mixture-of-experts ideas recur across the efficiency and agent work — but the record is too small to reveal a dominant core direction yet.
Collaboration36/100
With a collaboration score of 36, the visible network is compact: 5 distinct coauthors appear across the reviewed catalog. Zhenyu Zhang is the clear anchor, listed three times in the frequent-coauthor record, alongside Zhaowen Tao and Meng Chen. Yang holds a middle authorship position on the efficiency paper and the final position on the other two, none flagged as corresponding author.
Activity—
Activity data is insufficient: the catalog holds only three 2026 arXiv preprints and no longitudinal publication record, so cadence and recency cannot be scored. This axis would measure publishing rhythm and recent-year momentum across a longer span.
Reproducibility55/100
Reproducibility scores 55, the mean across 3 reviewed papers, each rated medium reproducibility readiness. Evolve's verdicts converge on the same reporting gaps: the L0-MoE work lacks variance analysis and MoE baselines, MOSCOPT reports no significance tests or cost-matched comparison, and HintMiner shows internal inconsistencies and undeclared baselines — methodologically interesting work whose empirical reporting is not yet audit-grade.
Research summary
Jiudong Yang is a machine-learning researcher whose reviewed work clusters around three linked problems: making dense large language models cheaper to serve, optimizing LLM agents directly in text space, and applying language models to developer Q&A content. The Evolve catalog contains three 2026 arXiv preprints. On efficiency, Yang contributed as a middle-listed author to L0-MoE, which uses L0 regularization to convert dense LLMs into sparse mixture-of-experts models and claims 2–2.5× acceleration at near-parity accuracy; Evolve finds the approach credible but notes missing variance analysis and absent MoE baselines. As last author of MOSCOPT, Yang proposes text-native joint optimization of a skill pool and its gating for LLM agents, reported to beat all baselines without significance tests or cost-matched comparison. HintMiner, also last-authored, mines question hints from Q&A web posts via self-supervised learning with a copy mechanism and BERT, reporting BLEU 36.17, though reviewers flag internal inconsistencies and undeclared baselines. Radar metrics show a split focus (score 0; three fields at roughly 33% each), a compact collaboration network (score 36; 5 distinct coauthors), and reproducibility at 55 across 3 papers. Output, impact, and activity remain unscored for lack of data.
Collaboration pattern
Yang's visible collaboration footprint is compact: 5 distinct coauthors across the reviewed catalog, with Zhenyu Zhang the clear anchor — appearing three times in the frequent-coauthor record — plus Zhaowen Tao and Meng Chen. Yang holds a middle authorship position on the efficiency paper and the final position on the agent and hint-mining papers, none flagged as corresponding. This pattern suggests a small, stable working group rather than a wide cross-institutional network.
Research keywords
Highlights
L0-MoE: pruning dense LLMs into sparse experts for 2–2.5× speedup
MOSCOPT: text-space joint optimization of skill pools and gating for LLM agents
HintMiner: self-supervised mining of question hints from Q&A posts (BLEU 36.17)
Timeline
- 2026Last-authored MOSCOPT on arXiv, a mixture-of-skills collective optimization method that jointly tunes a skill pool and gating for LLM agents in text space.
- 2026Last-authored HintMiner on arXiv, mining automatic question hints from Q&A web posts via self-supervised learning, with a reported BLEU of 36.17.
- 2026Middle-listed co-author of L0-MoE on arXiv, converting dense LLMs into sparse experts via L0 regularization with claimed 2–2.5× acceleration at near-parity accuracy.
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 scores and the field distribution are computed from a narrow bibliometric snapshot (3 reviewed 2026 arXiv preprints) and may not reflect the full publication record.
- Institution, country, and ORCID are absent from the source record; this profile is name-resolved from public bibliographic data and same-name authors may have been merged or split.
- Output, impact, and activity dimensions are unscored because citation counts and longitudinal productivity data are not present in the input.
- All reviewed papers are arXiv preprints dated 2026; no peer-reviewed venue information is available in the snapshot.