Tsinghua-affiliated researcher on LLM post-training: multi-teacher and data-free on-policy 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—
The output axis measures publication volume and its trend over time. This snapshot lacks a usable output score: the per-year work record contains only 2026 (2 works), so cumulative productivity cannot be assessed yet.
Impact—
The impact axis measures citation-based influence (h-index, i10-index, citations). Recorded values are currently 0 and 0, and the two reviewed 2026 preprints are too recent to have accrued citations, so no score is assigned yet.
Focus0/100
The radar records a focus score of 0. The field distribution is in fact narrow — 0.5 for machine learning (LLM post-training) and 0.5 for NLP (LLM post-training/knowledge distillation) — but the OpenAlex topic record spans four topics: Topic Modeling (2), Domain Adaptation and Few-Shot Learning (2), Natural Language Processing Techniques (1), and Reinforcement Learning in Robotics (1).
Collaboration59/100
Collaboration scores 59. The recent window records 18 distinct coauthors, and the frequent list flags repeat collaborators Gengsheng Li, Mingyang Song, and Jie Sun, alongside Mao Zheng, Ruiqi Liu, Qiyong Zhong, Haiyun Guo, Junfeng Fang, and Jinqiao Wang. Both reviewed papers carry middle-author, non-corresponding positions, consistent with team-based LLM post-training work.
Activity32/100
Activity scores 32. The record shows recent3yWorks = 2 and a countsByYear entry only for 2026 (2 works); both reviewed papers are 2026 arXiv preprints. Recorded publication volume in the last three years is low, so this is an early-stage activity profile.
Reproducibility55/100
Reproducibility scores 55 (reproMean 55 computed over 2 papers). Both reviewed papers carry medium reproReadiness: TrustMOPD reports strong gains across 5 seeds but no significance tests, and DF-OPD is well-controlled while its 0.13-point headline gap lacks error bars.
Research summary
Zeyuan Liu is a Tsinghua University-affiliated researcher whose recorded output centers on large language model post-training, specifically knowledge distillation. This profile is name-resolved from an OpenAlex-recorded snapshot (identity: openalex; no ORCID linked), which attributes two 2026 arXiv preprints reviewed by Evolve, on both of which Zeyuan Liu is a middle author and not corresponding author. The first, 'Distill What You Trust: Reliability-Aware Multi-Teacher On-Policy Distillation', presents TrustMOPD, a label-free token-level teacher-weighting scheme that reads reliability from RL displacement relative to a shared reference; Evolve notes strong gains across 5 seeds but no significance tests. The second, 'Data-free On-policy Distillation' (DF-OPD), argues that on-policy distillation barely needs its data and that a teacher can write its own prompts — provocative and well-controlled, though the 0.13-point headline gap lacks error bars. The field distribution splits evenly (0.5/0.5) between machine learning (LLM post-training) and NLP (LLM post-training/knowledge distillation), while the topic record also lists Topic Modeling and Domain Adaptation and Few-Shot Learning (2 each), plus Natural Language Processing Techniques and Reinforcement Learning in Robotics (1 each). Collaboration is broad: 18 distinct coauthors. Citation indices read 0 (h-index 0, i10-index 0), consistent with very recent output. Radar scores: collaboration 59, reproducibility 55, activity 32, focus 0; output and impact axes lack sufficient data.
Collaboration pattern
Zeyuan Liu collaborates within sizeable LLM post-training teams: the recent window records 18 distinct coauthors, and the frequent-collaborator list repeats Gengsheng Li, Mingyang Song, and Jie Sun, alongside Mao Zheng, Ruiqi Liu, Qiyong Zhong, Haiyun Guo, Junfeng Fang, and Jinqiao Wang. On both reviewed 2026 arXiv papers the position is middle author and not corresponding, indicating a contributor role within group projects rather than sole leadership.
Research keywords
Highlights
TrustMOPD: label-free, reliability-aware multi-teacher on-policy distillation
DF-OPD: data-free on-policy distillation
Consistent emphasis on controlled evaluation, with gaps in statistical rigor
Across both reviewed papers, evaluation uses multiple seeds or control conditions, yet neither reports significance tests or error bars. This pattern underlies Evolve's reproducibility mean of 55 over 2 papers, both rated medium reproReadiness.
Timeline
- 2026'Distill What You Trust: Reliability-Aware Multi-Teacher On-Policy Distillation' appears on arXiv (middle author, not corresponding), introducing TrustMOPD's label-free token-level teacher weighting.
- 2026'Data-free On-policy Distillation' appears on arXiv (middle author, not corresponding), with DF-OPD claiming OPD barely needs its data and teachers can write their own prompts.
- 2026The OpenAlex per-year record counts 2 works, both in 2026, both rated medium reproReadiness by Evolve; citation indices remain at 0.
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 are computed from an OpenAlex-recorded snapshot (identity: openalex); no ORCID is linked, and Zeyuan Liu is a common name, so this profile is name-resolved from public bibliographic data.
- Citation indices are currently 0 (h-index 0, i10-index 0) and the only counted year is 2026; both reviewed papers are very recent arXiv preprints, so impact cannot yet be judged.
- The OpenAlex topic record includes areas (topic modeling, domain adaptation and few-shot learning, reinforcement learning in robotics) that are not covered by the two reviewed papers.
- The reproducibility score (mean 55) is an Evolve assessment of 2 papers, not a community replication record.