Zhejiang University researcher in reinforcement learning policy optimization for LLM agents and multimodal models
3
papers here
2
reviewed by Evolve
2
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 is unscored: the OpenAlex snapshot carries no reliable publication-volume series beyond countsByYear (one work in 2023, four in 2026), which is too thin to rate productivity. This axis would measure total and annual output across venues, journals, and preprints.
Impact—
Impact is unscored: the recorded h-index and i10-index are both 0 and citation coverage is insufficient to judge reception of the work. This axis would measure citation counts, citing works, and influence on follow-up research.
Focus0/100
Focus scores 0/100, reflecting a dispersed OpenAlex topic profile: reinforcement learning in robotics (3 works), topic modeling (2), AI-based problem solving and planning (2), multimodal machine learning applications (2), and COVID-19 epidemiological studies (1). The field distribution splits evenly (0.5/0.5) between two adjacent ML subfields, though the Evolve-reviewed papers themselves cluster tightly on agent and multimodal policy optimization.
Collaboration60/100
Collaboration scores 60/100. The snapshot lists 19 distinct coauthors, with a recurring core — Jintao Chen, Jingxiao Yang, and Xuhong Zhang each appear twice in the frequent-coauthor list — alongside Wangjie Gan, Miao Pan, Fan Wang, Weiming Liu, Cheng Tan, and Wenqi Zhang. All three catalog papers are multi-authored, with first-author leadership on two and no corresponding-author role recorded.
Activity47/100
Activity scores 47/100. countsByYear shows one work in 2023 and four in 2026, with recent3yWorks set at 4; all three Evolve-reviewed papers are 2026 arXiv preprints. Output is therefore recent and concentrated in 2026 rather than sustained across many years, consistent with an early-stage publication record.
Reproducibility55/100
Reproducibility scores 55/100, a reproMean computed across 2 repro-scored papers, and all three catalog entries carry a medium reproReadiness rating. Evolve's verdicts temper the score: DEEPO lacks a code release and contains cross-table number inconsistencies, BATON has thin statistics and theory, and TASPO is undercut by draft artifacts despite strong ablations.
Research summary
Yingxuan Zhuang is a machine-learning researcher at Zhejiang University whose Evolve-reviewed work centers on reinforcement-learning policy optimization for large language models. The catalog holds three 2026 arXiv preprints. DEEPO, first-authored, couples semantic entropy and Rényi entropy within GRPO-style training to mitigate hallucination in multimodal LLMs; Evolve calls its evidence unusually rigorous but flags cross-table number inconsistencies and no code release. BATON, also first-authored, reweights token losses through Bayesian feedback attribution and trajectory mass normalization, delivering consistent gains across GRPO and GiGPO settings, though its statistics and theory stay thin. TASPO, a middle-authored contribution, reconciles privileged action-level process supervision with outcome-based credit, reported at +10.6% average across nine agent-RL settings with strong ablations undermined by draft artifacts. Credit assignment and hallucination mitigation are the connective tissue. The bibliometric snapshot is early-stage: OpenAlex records h-index and i10-index of zero, one work in 2023 and four in 2026 (four in the last three years), nineteen distinct coauthors, and topics spanning robotics RL, topic modeling, planning, and multimodal applications. Reproducibility averages 55 across two repro-scored papers, with every catalog entry rated medium.
Collaboration pattern
The collaboration pattern combines a recurring core team with a broader rotating pool: 19 distinct coauthors are recorded, and Jintao Chen, Jingxiao Yang, and Xuhong Zhang each appear twice in the frequent-coauthor list, alongside Wangjie Gan, Miao Pan, Fan Wang, Weiming Liu, Cheng Tan, and Wenqi Zhang. All three catalog papers are multi-authored; Zhuang leads two as first author while holding no corresponding-author role on any reviewed paper.
Research keywords
Highlights
Dual-entropy RL to curb hallucination in multimodal LLMs
Bayesian feedback attribution for LLM-agent policy optimization
Privileged process supervision reconciled with outcome-based credit
Timeline
- 2023Earliest point in the bibliometric record: OpenAlex countsByYear lists one publication this year, with no citation impact recorded for the period.
- 2026First-authored DEEPO appears on arXiv, proposing dual-entropy (semantic and Renyi) enhanced policy optimization to mitigate hallucination in multimodal LLMs.
- 2026First-authored BATON introduces Bayesian feedback attribution and trajectory mass normalization for LLM-agent policy optimization, with consistent gains over GRPO/GiGPO baselines.
- 2026Contributes as a middle author to TASPO, which reweights GRPO credit with privileged process supervision and reports +10.6% average improvement across nine agent-RL settings.
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 are drawn from OpenAlex only; no ORCID is provided and the profile is name-resolved from public bibliographic data, so records for 'Yingxuan Zhuang' may merge or split.
- Citation-impact coverage is insufficient: h-index and i10-index are recorded as 0, and the output and impact dimensions are unscored (dimMissing) rather than zero-rated on merit.
- All three reviewed papers are 2026 arXiv preprints; the catalog contains no peer-reviewed venue information.
- The OpenAlex topic list includes adjacent areas (topic modeling, COVID-19 epidemiological studies) that may reflect name-resolution noise rather than this researcher's current direction.