Alibaba Group researcher advancing LLM post-training: counterfactual on-policy distillation for instruction following
2
papers here
1
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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—
Output data is insufficient to score this axis. The snapshot only captures 6 works in 2026 with no aggregated publication total, venue record, or output typology. This axis would measure total publication volume and productivity over time.
Impact—
Impact cannot be scored yet: the snapshot reports h-index 0 and i10-index 0 with no citation counts available. This axis would measure citations, citation velocity, and uptake of the researcher's work by the wider community.
Focus100/100
Focus scores 100/100, the maximum concentration. The assigned field is machine learning (LLM post-training), and the OpenAlex topic distribution clusters tightly: Topic Modeling leads with 6, followed by Natural Language Processing Techniques (3) and Reinforcement Learning in Robotics (3), with Information Retrieval (2) and Domain Adaptation/Few-Shot Learning (2) trailing. The profile reads as a deliberate LLM-training specialization with a small applied periphery.
Collaboration55/100
The collaboration axis scores 55/100. The snapshot records 15 distinct coauthors, with Tianze Xu and Chao Ma appearing twice in the frequent-coauthor list, marking the closest repeated ties. Both catalog papers are multi-author team efforts — a first-author position on CC-OPD and a middle position on PAC — a pattern consistent with large industry research groups at Alibaba rather than solo work.
Activity57/100
Activity scores 57/100. The counts-by-year record shows 6 works in 2026 and 6 works over the last three years, meaning all visible output is concentrated in a single recent year. This is an early-stage publication rhythm — intense current activity with no longer publication history yet captured in the snapshot to establish a sustained trend.
Reproducibility55/100
Reproducibility scores 55/100, computed from a mean over 1 reviewed paper. The reviewed study, CC-OPD, carries a medium repro-readiness rating: Evolve notes it achieves the best average across 7 benchmarks over 2 Qwen model pairs, but flags single-seed runs and no confidence intervals, so reported gains should be interpreted with caution until multi-seed evidence appears.
Research summary
Yanzhao Zheng is a researcher at Alibaba Group (China) whose visible work centers on machine learning for LLM post-training, with the OpenAlex-assigned field recorded as machine learning (LLM post-training). The bibliometric snapshot is early-stage: counts-by-year data show 6 works in 2026 — the only year with output — and recent3yWorks also equals 6, while h-index and i10-index both read 0. The Evolve catalog includes two 2026 arXiv papers. The first is CC-OPD (Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following), where Zheng is first author: the method shapes on-policy distillation rewards through leave-one-out teacher ablation to isolate each constraint's contribution, achieving the best average across 7 benchmarks over 2 Qwen model pairs, though Evolve flags single-seed runs and missing confidence intervals and rates reproducibility readiness as medium. The second, PAC (Progress-Augmented Advantage Curriculum for Multi-Task Reinforcement Learning of LLMs), lists Zheng as a middle author and awaits full review. Topic distribution shows concentration in Topic Modeling (6), with Natural Language Processing Techniques (3), Reinforcement Learning in Robotics (3), Information Retrieval (2), and Domain Adaptation/Few-Shot Learning (2). Collaboration is team-heavy: 15 distinct coauthors, with Tianze Xu and Chao Ma recurring. Radar scores: focus 100, activity 57, collaboration 55, reproducibility 55; output and impact axes lack sufficient data.
Collaboration pattern
Yanzhao Zheng works in a visibly team-based industry setting: the snapshot lists 15 distinct coauthors, with Tianze Xu and Chao Ma each appearing twice in the frequent-coauthor list — the strongest repeated ties. Both catalog papers are multi-author efforts (first author on CC-OPD, middle on PAC), consistent with Alibaba Group's large LLM post-training teams. Other recurring collaborators include Yuanqiang Yu, Hangcheng Zhu, Ruohui Huang, Zhentao Zhang, Jihuai Zhu, Jiashun Liu, Xinle Deng, and Baohua Dong.
Research keywords
Highlights
Counterfactual, leave-one-out reward shaping for multi-constraint instruction following
Progress-augmented advantage curriculum for multi-task RL of LLMs
A 2026 arXiv co-authorship (middle author position) on PAC, which proposes a progress-augmented advantage curriculum for multi-task reinforcement learning of large language models; this paper is in the catalog but has not yet received a full Evolve review.
Topic footprint spanning LLM training and applied machine learning
The OpenAlex topic distribution shows Topic Modeling (6), Natural Language Processing Techniques (3), Reinforcement Learning in Robotics (3), Information Retrieval and Search Behavior (2), and Domain Adaptation and Few-Shot Learning (2), indicating a concentrated LLM-training core with applied ML breadth.
Timeline
- 2026First-author arXiv paper CC-OPD (Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following) appears, with a full Evolve review.
- 2026Co-authors PAC (Progress-Augmented Advantage Curriculum for Multi-Task Reinforcement Learning of LLMs) in a middle author position on arXiv.
- 2026Counts-by-year data record 6 works in 2026 — the only year with output and equal to the 3-year total, marking an early-stage burst of activity.
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.
- All bibliometrics come from OpenAlex (identity: openalex); no ORCID is linked, so this profile is name-resolved from public bibliographic data and may be affected by author disambiguation.
- Citation impact indicators (h-index 0, i10-index 0) come from the same snapshot and likely reflect an early-stage record or incomplete name disambiguation rather than measured influence.
- Only 1 of the 2 catalog papers (CC-OPD) has a full Evolve review with verdict, keywords, and repro-readiness; the PAC paper is listed without review data.
- The output and impact radar axes are missing in this snapshot, so scores are shown for four of six dimensions only.