Tencent researcher in LLM post-training, centered on on-policy and multi-teacher knowledge distillation.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
Not scored: per-work output volume and venue-quality data are insufficient in this snapshot. This axis would measure publication count, venue tier, and preprint-to-peer-review conversion over the recent window.
Impact—
Not scored: citation-based impact cannot be assessed here — OpenAlex records h-index 0 and i10-index 0 for this profile, consistent with incomplete citation capture rather than evidence of no influence. This axis would measure citations, uptake, and downstream reuse.
Focus0/100
Focus scores 0 in this snapshot: the OpenAlex topic distribution is spread across five areas — Topic Modeling (19), NLP Techniques (12), RL in Robotics (8), Domain Adaptation and Few-Shot Learning (8), and Speech Recognition and Synthesis (4). The Evolve catalog narrows this: both reviewed works sit in LLM post-training, with fieldDist split evenly between machine learning (LLM post-training) and NLP (knowledge distillation).
Collaboration59/100
Collaboration scores 59: 18 distinct co-authors across 22 works in the last three years, a mid-range breadth for an industry researcher. The recurring names — Gengsheng Li, Mingyang Song, and Jie Sun, each appearing twice in the frequent-coauthor list — suggest a stable core team, with the remaining partners rotating across projects.
Activity91/100
Activity scores 91, the strongest axis: 22 works in the last three years, with by-year counts ramping sharply from 2 in 2025 to 20 in 2026. This cadence matches an applied lab tempo — rapid iteration on distillation methods released as arXiv preprints, two of which Evolve reviewed in 2026.
Reproducibility55/100
Reproducibility scores 55, averaged over the 2 papers Evolve assessed; both carry a medium repro-readiness rating. The reviewed work reports multi-seed evidence — TrustMOPD across 5 seeds — but neither paper includes significance tests or error bars, which caps the rating despite generally well-controlled experiments.
Research summary
Mao Zheng is a researcher at Tencent (China) whose recorded work concentrates on post-training for large language models, specifically on-policy knowledge distillation. The profile is resolved from OpenAlex bibliographic data (no linked ORCID), and its bibliometric window is short: OpenAlex counts 2 works in 2025 rising to 20 in 2026, and 22 works over the last three years, which drives a very high activity score (91). Topic coverage is broad — Topic Modeling (19), NLP Techniques (12), Reinforcement Learning in Robotics (8), Domain Adaptation and Few-Shot Learning (8), and Speech Recognition and Synthesis (4) — yet the Evolve-reviewed record is tightly focused: two 2026 arXiv papers, both on multi-teacher on-policy distillation, with the field distribution split evenly between machine learning (LLM post-training) and NLP (knowledge distillation). The first, TrustMOPD, introduces label-free token-level teacher weighting via RL displacement from a shared reference, with gains across 5 seeds but no significance tests. The second, DF-OPD, argues on-policy distillation barely needs its training data and that a teacher can write its own prompts, with well-controlled but error-bar-free headline gaps (0.13 points). Both papers list Mao Zheng as a middle, non-corresponding author, and citation aggregates (h-index 0, i10-index 0) are not yet meaningful.
Collaboration pattern
With 18 distinct co-authors across 22 recent works (collaboration score 59), Mao Zheng works in a mid-sized, stable team. Gengsheng Li, Mingyang Song, and Jie Sun recur in the frequent-coauthor list, indicating repeated collaboration on the same line of work. Both Evolve-reviewed papers list Mao Zheng as a middle author, not corresponding — consistent with a contributing role inside a larger Tencent group rather than sole leadership.
Research keywords
Highlights
TrustMOPD: reliability-aware multi-teacher on-policy distillation
DF-OPD: on-policy distillation with almost no data
A sharp 2026 ramp around one distillation question
Recorded works jump from 2 in 2025 to 20 in 2026, and both reviewed papers attack the same problem — which teacher signals to trust in on-policy distillation — from two angles: reliability-aware weighting and data-free training.
Timeline
- 2025Earliest activity visible in this snapshot: OpenAlex records 2 works in 2025, the first year with any count.
- 2026Recorded output jumps to 20 works in 2026, concentrated in LLM post-training and knowledge distillation.
- 2026Evolve reviews 'Distill What You Trust' (arXiv): TrustMOPD adds label-free token-level teacher weighting to multi-teacher on-policy distillation, with gains across 5 seeds but no significance tests.
- 2026Evolve reviews 'Data-free On-policy Distillation' (arXiv): DF-OPD shows OPD barely needs its data and a teacher can write its own prompts; well-controlled but the 0.13-point gap lacks error bars.
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.
- Bibliometrics come from OpenAlex only; no ORCID is linked, and 'Mao Zheng' is a common name, so this profile is name-resolved from public bibliographic data and may merge or split identities.
- OpenAlex records h-index 0 and i10-index 0; the output and impact dimensions are therefore marked missing rather than scored.
- Only 2 papers underlie the reviewed evidence — both 2026 arXiv preprints with middle, non-corresponding authorship — so the reproducibility mean (55) reflects just those two assessments.
- No venue, affiliation-history, or career data beyond 'Tencent (China)' is available; education, awards, and prior roles are unknown.