Baidu ML researcher on efficient MoE inference and rubric-based RL post-training 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 (2026-10-09) 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 publication-volume data is available yet: total output, venue mix, and yearly counts are not in the current bibliometric records. This axis would measure aggregate research productivity over time, including how output is distributed across venues and authorship positions.
Impact—
Citation counts, h-index-style indicators, and per-paper influence are absent from the current records, so impact is unscored. This axis would measure how widely these inference-efficiency and post-training methods are cited, adopted, and built upon by other researchers.
Focus0/100
Scores 0, grounded in the raw field distribution: efficient LLM inference and LLM post-training / RLHF each account for exactly 50% of the cataloged output, and the two 2026 papers sit one per subfield. Current evidence therefore shows breadth across two adjacent LLM directions rather than concentration in a single line — a reading that may shift as more papers enter the catalog.
Collaboration62/100
Scores 62 on a 0-100 scale, computed from real bibliometric data: 21 distinct coauthors are recorded, and Peisong Wang, Yuantian Shao, and Yuanteng Chen each appear repeatedly in the frequent-coauthor list, pointing to a stable core team inside Baidu. Combined with middle and last authorship positions on the two cataloged papers, this suggests an established contributor embedded in a persistent industry research group.
Activity—
Recent-year activity metrics — papers per year and submission cadence — are not available in the current data. This axis would measure ongoing output tempo and whether research activity is accelerating or declining over recent years.
Reproducibility38/100
Scores 38, a reproMean computed over 2 scored papers (reproPapers = 2). The MoE pruning study is rated medium readiness — it states explicit retention ratios and reports that keeping 2/3 of chosen experts preserves 98.8% of performance — while RLVR² is rated low: single-run experiments, no significance tests, and closed evaluation data currently limit independent replication of its reward-aggregation results.
Research summary
Jing Liu is a machine learning researcher at Baidu (China) whose cataloged work centers on large language models, split evenly — 50% and 50% in the raw field distribution — between efficient LLM inference and LLM post-training / RLHF. Two 2026 arXiv papers anchor this full-tier profile. The first, an empirical study of dynamic expert pruning in fine-grained mixture-of-experts LLMs with a middle authorship position, shows that uniformly keeping 2/3 of the chosen experts preserves 98.8% of MoE performance and that dynamic pruning pays off only under aggressive budgets; Evolve rated it convincing, solid work with medium reproducibility readiness. The second, RLVR², with a last authorship position, turns rubric scores into within-group Bradley-Terry rankings used as reinforcement learning reward, reporting broad gains across three model scales — though results are single-run, lack significance tests, and rest on closed data, giving low reproducibility readiness. Collaboration signals record 21 distinct coauthors (score 62) with Peisong Wang, Yuantian Shao, and Yuanteng Chen recurring across papers; the focus axis scores 0, reflecting the even two-way split. Since 'Jing Liu' is a common name, this profile is name-resolved from public bibliographic data.
Collaboration pattern
Jing Liu's network spans 21 distinct coauthors (collaboration score 62). Peisong Wang, Yuantian Shao, and Yuanteng Chen appear repeatedly in the frequent-coauthor list, marking them as the closest recurring collaborators, while the remaining names indicate more project-specific partnerships. On the two cataloged papers Jing Liu holds middle and last authorship positions, neither corresponding — a pattern consistent with a senior or core contributing member of a Baidu research team rather than a first-author-led line.
Research keywords
Highlights
Keeping 2/3 of chosen experts preserves 98.8% of MoE performance
RLVR² converts rubric scores into Bradley-Terry ranking rewards
Timeline
- 2026Co-authored (middle position) the empirical study of dynamic expert pruning in fine-grained MoE LLMs, released on arXiv.
- 2026Co-authored (last position) RLVR², on rubric-based Bradley-Terry ranking rewards for LLM post-training, released on arXiv.
- 2026Released both cataloged papers in the same year, splitting output evenly between efficient inference and post-training/RLHF.
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 signals (coauthor counts, field distribution, radar scores) derive from public bibliometric data only; no ORCID is linked, and 'Jing Liu' is a very common name, so this profile is name-resolved from public bibliographic records.
- The catalog covers only two 2026 arXiv preprints; peer-review status is not recorded, and no citation or impact metrics are available.
- Reproducibility and focus scores rest on just two scored papers (reproPapers = 2), so radar readings should be treated as early-stage evidence.
- Output, impact, and activity axes are unscored because the underlying bibliometric data is insufficient, not because the researcher lacks output.