Ant Group researcher spanning LLM post-training efficiency and world-model-based VLA reinforcement learning
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) 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 snapshot has no publication counts, venue history, or per-year totals beyond the two 2026 arXiv preprints in the Evolve catalog. This axis would measure total publication volume, output cadence, and authorship roles across venues.
Impact—
Not yet scored: no citation counts, h-index, or downstream-usage signals were provided. This axis would measure how often the two cataloged papers, and any uncataloged work, are cited, adopted, or built upon by other researchers.
Focus0/100
Scores 0 on focus, reflecting an evenly split field distribution: the two cataloged papers divide 0.5/0.5 between machine learning (large language models) and robot learning / reinforcement learning. Rather than one deep vertical, the recorded portfolio spreads across two distinct areas — sparse on-policy distillation for LLMs and world-model-based VLA policy optimization — linked mainly by a shared efficiency motive. With only two papers, the split may partly reflect sample size.
Collaboration56/100
Scores 56 with 16 distinct coauthors recorded across the two cataloged papers. On both, Haonan Wang appears in a middle, non-corresponding position, consistent with contributing specific components inside larger Ant Group teams. The frequent-coauthor list includes Huanxin Sheng, Zhiling Ye, Jian Wang, Jinjie Gu, Jian Kang, Haoxiang Ren, Zhilong Zhang, Runjie Xu, Yihao Sun, Yifei Sheng, and Nan Tang, indicating a stable internal group rather than loose external ties.
Activity—
Not yet scored: the input contains no multi-year publication timeline — only two 2026 arXiv preprints. This axis would measure publishing cadence and consistency over recent years, which cannot be inferred from a single-year snapshot.
Reproducibility55/100
Averages 55 across the two rated papers, each rated medium reproducibility readiness. Evolve's reviews flag the limitations behind that score: the on-policy distillation paper reports many gains within its own noise bounds, and the U-GROW paper shows consistent sim/real improvements over VLA-MBPO but from single-run results without error bars. Theory and experiment coverage is broad in both cases, yet no independent replication is recorded in the catalog.
Research summary
Haonan Wang is a researcher at Ant Group (China) whose cataloged work spans two adjacent efficiency problems in modern machine learning: making large-language-model post-training cheaper and making model-based reinforcement learning for robot policies more sample-efficient. The Evolve catalog holds two 2026 arXiv preprints with a middle, non-corresponding authorship position on each. The first, '1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation', proposes IER for selecting reliable-gradient tokens in sparse on-policy distillation and claims that 1% of tokens can suffice; Evolve's review credits broad theory and experiments while cautioning that many reported gains fall within noise. The second, 'Prioritized Rollouts for Efficient World Model-based Vision-Language-Action Policy Optimization', introduces U-GROW, uncertainty-guided rollout-start sampling for world-model-based VLA training, with consistent simulation and real-robot gains over VLA-MBPO, though from single-run results without error bars. Bibliometrically, the recorded field distribution is split evenly (0.5/0.5) between machine learning (large language models) and robot learning / reinforcement learning, the collaboration dimension scores 56 with 16 distinct coauthors, and reproducibility averages 55 across the two rated papers. Both carry a medium reproducibility-readiness rating, shaping a profile of a technically broad, efficiency-driven contributor whose strongest claims still await independent verification.
Collaboration pattern
Both cataloged papers list Haonan Wang in a middle, non-corresponding position, consistent with contributing specific components inside larger Ant Group research teams. Sixteen distinct coauthors are recorded across the two papers, and the frequent-coauthor list — including Huanxin Sheng, Zhiling Ye, Jian Wang, Jinjie Gu, Jian Kang, Haoxiang Ren, Zhilong Zhang, Runjie Xu, Yihao Sun, Yifei Sheng, and Nan Tang — suggests a stable internal group spanning both the LLM and robot-learning tracks rather than loose external partnerships.
Research keywords
Highlights
1% of tokens argued sufficient for on-policy distillation
U-GROW: uncertainty-guided rollouts for world-model VLA training
Timeline
- 2026Cataloged arXiv preprint on gradient estimation in on-policy distillation ('1% of Tokens Can Be Enough'), middle authorship, non-corresponding.
- 2026Cataloged arXiv preprint proposing U-GROW for world-model-based vision-language-action policy optimization, middle authorship, non-corresponding.
- 2026Recorded field distribution lands at an even 0.5/0.5 split between large-language-model work and robot learning / reinforcement learning, with 16 distinct coauthors across the two papers.
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.
- Dimension scores are computed from a public bibliometric snapshot; no citation counts, h-index, or venue-tier data were included in the input.
- 'Haonan Wang' is a common name; this profile is name-resolved from public bibliographic data, and no ORCID was available to confirm identity.
- Only two papers are in the Evolve catalog, both 2026 arXiv preprints; career history, awards, and affiliation timeline were not provided.
- Output, impact, and activity axes are unscored, and the reproducibility mean of 55 averages just two papers rated medium readiness.