Alibaba researcher in LLM post-training: reinforcement learning for instruction following and multi-task LLM training
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
1
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—
Not yet scored: the OpenAlex record is too early for reliable output metrics, with h-index 1 and i10-index 0. This axis would measure publication volume and productivity trajectory once a fuller work and citation history accumulates.
Impact—
Not yet scored: citation-based influence cannot be assessed from a record with h-index 1, i10-index 0, and no aggregate citation totals. This axis would measure citation uptake, downstream adoption, and scholarly reach as the record matures.
Focus100/100
Focus scores 100: OpenAlex's field distribution assigns the entire indexed record to a single field, machine learning (LLM post-training), and both 2026 Evolve-catalog papers pursue the same theme of reinforcement learning for LLM training — instruction following via on-policy distillation and multi-task curriculum RL. The broader topic list (topic modeling, RL in robotics, NLP, few-shot learning) likely reflects tagging noise rather than a second agenda.
Collaboration55/100
Collaboration scores 55, from 15 distinct coauthors in the snapshot. Recurring partners include Tianze Xu and Chao Ma (each listed twice among frequent coauthors), alongside Yuanqiang Yu, Yanzhao Zheng, Ruohui Huang, Zhentao Zhang, Jihuai Zhu, Jiashun Liu, Xinle Deng and Baohua Dong. On both catalog papers Zhu holds a middle author position without corresponding status, consistent with contributing-member roles inside a sizeable Alibaba industry team.
Activity75/100
Activity scores 75, driven by the OpenAlex counts-by-year curve: 2 works indexed in 2025 rising to 10 in 2026, with 12 works over the recent three-year window. Nearly all indexed output lands in a single year, a steep early-career ramp centered on 2026 arXiv preprints rather than a long, steady publication cadence.
Reproducibility55/100
Reproducibility scores 55, computed from a single Evolve-reviewed paper (mean 55, medium readiness). For CC-OPD, Evolve notes the method shapes on-policy distillation rewards via leave-one-out teacher ablation and achieves the best average across 7 benchmarks on 2 Qwen model pairs, but flags single-seed runs and missing confidence intervals. With only one reviewed paper, this axis carries low evidential weight.
Research summary
Hangcheng Zhu is an early-career researcher at Alibaba Group (China) whose indexed work centers on machine learning for LLM post-training. The profile is name-resolved from OpenAlex bibliographic data (ORCID 0009-0003-0748-612X). The record is young but accelerating: 2 works indexed in 2025 versus 10 in 2026, 12 works over the recent three-year window, with an h-index of 1 and an i10-index of 0 — typical of a first- or second-year publishing record. OpenAlex assigns the entire field distribution to a single field, machine learning (LLM post-training), yielding a maximal focus score. Two 2026 arXiv papers appear in the Evolve catalog, both with middle author positions and no corresponding role: Counterfactual Constraint-Conditioned On-Policy Distillation for Multi-Constraint Instruction Following (CC-OPD), which shapes on-policy distillation rewards via leave-one-out teacher ablation and reports the best average across 7 benchmarks on 2 Qwen model pairs, and PAC: Progress-Augmented Advantage Curriculum for Multi-Task Reinforcement Learning of LLMs. Evolve's reproducibility review covers only the first paper (mean 55, medium readiness), flagging single-seed runs without confidence intervals. Broader OpenAlex topic assignments — topic modeling, reinforcement learning in robotics, NLP techniques, domain adaptation and few-shot learning — likely reflect disambiguation noise in automated topic tagging rather than a confirmed second research agenda.
Collaboration pattern
Within the OpenAlex snapshot Hangcheng Zhu has 15 distinct coauthors; the frequent list is topped by Tianze Xu and Chao Ma, each appearing twice, followed by Yuanqiang Yu, Yanzhao Zheng, Ruohui Huang, Zhentao Zhang, Jihuai Zhu, Jiashun Liu, Xinle Deng and Baohua Dong. Both Evolve-catalog papers list Zhu as a middle author without corresponding status, suggesting a contributing specialist role inside a large Alibaba Group LLM team rather than solo or lead-author work.
Research keywords
Highlights
CC-OPD: counterfactual reward shaping for multi-constraint instruction following
PAC: progress-augmented curriculum for multi-task RL of LLMs
A second 2026 arXiv paper, with a middle author position, proposes Progress-Augmented Advantage Curriculum for multi-task reinforcement learning of LLMs; it is in the catalog without an Evolve verdict, keywords, or reproducibility assessment yet.
2026 output ramp: 10 indexed works in a single year
OpenAlex counts-by-year shows 2 works in 2025 rising to 10 in 2026 (12 over the recent three years), a steep early ramp concentrated in LLM post-training and reinforced by two arXiv preprints entering the Evolve catalog.
Timeline
- 2025First indexed works appear: OpenAlex counts 2 publications in 2025.
- 2026CC-OPD arXiv paper on counterfactual, constraint-conditioned on-policy distillation for multi-constraint instruction following (middle author, no corresponding role).
- 2026PAC arXiv paper on progress-augmented advantage curriculum for multi-task reinforcement learning of LLMs (middle author, no corresponding role).
- 2026OpenAlex indexes 10 works in 2026 (12 over the recent three-year window), marking the steepest output year to date.
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; identity is name-resolved from public bibliographic data tied to ORCID 0009-0003-0748-612X.
- Output and impact are unscored (insufficient data): h-index 1 and i10-index 0 with no aggregate citation totals reflect a very early record.
- Only one catalog paper has an Evolve review; PAC lacks a verdict, keywords, and reproducibility assessment, so the reproducibility score rests on a single paper.
- OpenAlex topic assignments (topic modeling, RL in robotics, IR and search behavior, few-shot learning) may include name-disambiguation noise and should not be read as confirmed research lines.