Ruohui Huang (Alibaba Group) — LLM post-training: on-policy distillation, counterfactual reward shaping, multi-task RL.
2
papers here
1
reviewed by Evolve
0
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—
The output axis is not yet scored: the snapshot lacks a verified total publication count or venue breakdown, so this axis would measure cumulative paper volume and its distribution across venues and years once author attribution is confirmed.
Impact—
Impact cannot be scored yet: citation-based measures are missing from the usable data — the snapshot's h-index and i10-index both read 0 — so this axis would measure citations, citation velocity, and uptake once citation attribution is reliable.
Focus100/100
Focus reaches the maximum score of 100 because the classified catalog output sits in a single field, machine learning (LLM post-training), with no competing field assignments in fieldDist. The recent papers — counterfactual on-policy distillation and a progress-augmented curriculum for multi-task RL — stay within LLM post-training. Broader OpenAlex topic counts (topic modeling 6, reinforcement learning in robotics 4, domain adaptation and few-shot learning 4) likely reflect namesake records rather than genuine dispersion.
Collaboration55/100
Collaboration scores 55, built on 15 distinct coauthors in the snapshot. The recurring-partner list centers on Yuanqiang Yu, Yanzhao Zheng, Hangcheng Zhu, Zhentao Zhang, Tianze Xu, and Chao Ma — with Tianze Xu and Chao Ma each appearing twice — suggesting a stable core team. Both catalog papers place Ruohui Huang in the last author position without a corresponding-author flag, consistent with a senior contributor inside one Alibaba Group team rather than a wide external network.
Activity64/100
Activity scores 64 on recent output volume: recent3yWorks counts 8 works, and countsByYear places all 8 in 2026 — a concentrated burst that includes both catalog papers. No works appear in intermediate years; the only other records fall in 2000 (1) and 2009 (1), likely namesake entries. Within the verifiable recent window, 8 works in a single year indicates a high-intensity, production-focused period.
Reproducibility55/100
Reproducibility scores 55 (reproMean), computed from a single evaluated paper (reproPapers 1), so confidence is limited. That paper, CC-OPD, carries a medium reproReadiness rating: Evolve's verdict reports the best average across 7 benchmarks over 2 Qwen model pairs, but flags single-seed runs and the absence of confidence intervals as reasons for caution, with no variance reporting behind the headline gains.
Research summary
Ruohui Huang is a researcher at Alibaba Group (China) whose verifiable recent output centers on post-training of large language models. The Evolve catalog lists two 2026 arXiv papers, both with Huang in the last author position: CC-OPD, a counterfactual constraint-conditioned on-policy distillation method for multi-constraint instruction following, and PAC, a progress-augmented advantage curriculum for multi-task reinforcement learning of LLMs. The bibliometric snapshot (OpenAlex) records 8 works in the last three years, all dated 2026, and 15 distinct coauthors, giving focus the maximum score of 100 (a single classified field: machine learning, LLM post-training), activity 64, and collaboration and reproducibility 55 each. The reproducibility figure rests on one evaluated paper: Evolve rates CC-OPD medium for reproducibility, crediting the strongest average across 7 benchmarks over 2 Qwen model pairs while cautioning that runs are single-seed with no confidence intervals. Interpretation requires care: the snapshot's h-index and i10-index read 0, output and impact axes are unscored, and topic counts spanning topic modeling, robotics reinforcement learning, and few-shot learning — plus stray 2000 and 2009 work records — likely reflect namesake entries rather than this researcher's actual range. Within the verified scope, this is an early public-record profile of a concentrated, team-embedded LLM post-training contributor.
Collaboration pattern
The collaboration pattern is a compact, team-internal network: 15 distinct coauthors, with the most frequent partners being Yuanqiang Yu, Yanzhao Zheng, Hangcheng Zhu, Zhentao Zhang, Tianze Xu, and Chao Ma (Tianze Xu and Chao Ma each recur twice in the frequency list). Both 2026 catalog papers place Ruohui Huang as last author without a corresponding-author flag, consistent with senior-team participation in an Alibaba Group LLM post-training group rather than broad external collaboration.
Research keywords
Highlights
Leave-one-out teacher ablation shapes on-policy distillation rewards
Best average across 7 benchmarks on 2 Qwen pairs — with caveats
Progress-augmented curriculum for multi-task RL of LLMs
The second catalog paper, PAC, proposes a progress-augmented advantage curriculum for multi-task reinforcement learning of LLMs; it is listed without an Evolve verdict, keywords, or reproducibility rating yet.
Timeline
- 2000Earliest work record under this name appears in the OpenAlex countsByYear (1 work); given the profile's concentrated 2026 activity, this record may belong to a namesake.
- 2009A single additional work record appears (countsByYear 2009: 1); its attribution to this researcher remains uncertain.
- 2026Output concentrates here: 8 works recorded in countsByYear (matching recent3yWorks 8), including both catalog papers — CC-OPD on multi-constraint instruction following and PAC on multi-task RL curricula — with Huang as last author on each.
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 derive from an OpenAlex snapshot only; the profile is name-resolved from public bibliographic data without an ORCID, so namesake conflation is possible (e.g., the 2000 and 2009 work records and topic counts spanning topic modeling and robotics RL).
- Output and impact dimensions are unscored: verified publication totals and citation attribution are unavailable; the snapshot reports h-index 0 and i10-index 0, likely reflecting incomplete citation matching.
- The reproducibility score (reproMean 55) rests on a single evaluated paper; the second catalog paper (PAC) has no Evolve verdict, keywords, or reproducibility rating.
- Profile scope is limited to OpenAlex bibliometrics plus papers reviewed by Evolve; no affiliation history, award, or funding data were used.