Alibaba researcher spanning LLM post-training and scientific knowledge infrastructure, with two 2026 arXiv papers.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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 current snapshot contains no publication-count or venue data, so output is unscored. This axis would measure the volume and cadence of the researcher's papers across their full publication record.
Impact—
No citation counts, h-index, or influence indicators are available yet, so impact is unscored. This axis would capture how often and how widely the researcher's work is cited by later studies.
Focus0/100
The focus score is 0, reflecting an even 50/50 split of cataloged work between two fields: machine learning for LLM post-training, and scientific information retrieval / AI for Science knowledge infrastructure. Both 2026 papers sit in different fields, so the profile reads as a deliberate two-track agenda rather than a single concentrated line; focus may rise if one track consolidates.
Collaboration68/100
Collaboration scores 68, based on 29 distinct co-authors in the underlying bibliometric data and a core group of 12 recurring collaborators — including Yanzhao Zheng, Yuanqiang Yu, Tianze Xu, and Zhentao Zhang. Both cataloged 2026 papers list the researcher in a middle author position, consistent with team-based industrial research rather than solo or corresponding-author work.
Activity—
Recent publication cadence and ongoing-research signals are missing from the data, leaving activity unscored. This axis would measure how consistently the researcher has published across recent years.
Reproducibility38/100
Reproducibility averages 38 across the 2 papers Evolve has reviewed: the counterfactual on-policy distillation work is rated medium readiness, while the Large Knowledge Model paper is rated low, with small untested margins and unmeasured graph fidelity. The pattern suggests capable engineering paired with reporting practices — single-seed runs, no confidence intervals — that currently limit independent verification.
Research summary
Chao Ma is a researcher at Alibaba Group whose cataloged work spans two tracks: machine learning for LLM post-training, and scientific information retrieval / AI for Science knowledge infrastructure. Evolve's catalog includes two 2026 arXiv papers, in both of which Ma holds a middle author position without corresponding-author status — a signature of contribution within large industrial research teams. The first, on counterfactual constraint-conditioned on-policy distillation (CC-OPD), shapes distillation rewards through leave-one-out teacher ablation and reports the best average score across 7 benchmarks over 2 Qwen model pairs; Evolve rates its reproducibility as medium, noting single-seed runs and no confidence intervals. The second, Large Knowledge Model, converts more than 40 million papers into reasoning graphs and reports question-answering gains up to 14.69 percentage points; Evolve rates it low on reproducibility, citing small untested margins and unmeasured graph fidelity. Radar scores place collaboration at 68 — grounded in 29 distinct co-authors and 12 recurring collaborators — and reproducibility at 38 averaged over the two reviewed papers, while focus registers 0 from an even 50/50 field split. Output, impact, and activity metrics are unavailable in the current snapshot and are left unscored. The profile is name-resolved from public bibliographic data; 'Chao Ma' is a common name.
Collaboration pattern
Ma's collaboration footprint includes 29 distinct co-authors, with a recurring core of 12 — Yanzhao Zheng, Yuanqiang Yu, Tianze Xu, Zhentao Zhang, Jihuai Zhu, Baohua Dong, and others — yielding a collaboration score of 68. Both cataloged 2026 papers carry middle author positions and no corresponding-author role, consistent with contributing-specialist work inside large Alibaba research teams rather than solo or lead-author projects.
Research keywords
Highlights
Counterfactual reward shaping lifts multi-constraint instruction following
Reasoning graphs over 40M+ papers for scientific QA
A two-track agenda bridging post-training and science infrastructure
The two cataloged 2026 papers split evenly between LLM post-training and AI-for-science knowledge infrastructure — a 50/50 field distribution that drives the 0 focus score while showing range across model training and knowledge systems.
Timeline
- 2026Publishes CC-OPD on arXiv: counterfactual constraint-conditioned on-policy distillation for multi-constraint instruction following, with middle-position authorship.
- 2026Co-authors Large Knowledge Model on arXiv, building reasoning graphs from 40M+ papers in the AI-for-science knowledge infrastructure track.
- 2026Both papers enter Evolve's catalog with reproducibility ratings of medium and low, establishing a full-tier profile spanning two fields with a 50/50 split.
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.
- Bibliometric scores are computed from the provided bibliometric snapshot only; publication counts, citations, and yearly activity are unavailable, so output, impact, and activity are unscored.
- 'Chao Ma' is a common name; this profile is name-resolved from public bibliographic data and may merge or split identities.
- The evidence base is two 2026 arXiv preprints reviewed by Evolve, both with middle author positions and no corresponding-author entries in the catalog.
- The reproducibility score of 38 rests on only 2 reviewed papers, too small a sample for a stable estimate.