Wangjie Gan
Zhejiang University of Science and Technology · China
GRPO-focused RL researcher tackling hallucination and credit assignment in multimodal and agentic LLMs
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—
Output cannot be scored: this snapshot lacks verified publication totals beyond the 2026 window, so career-wide volume and consistency are unknown. This axis would measure annual peer-reviewed production and how steadily that output is sustained over time.
Impact—
Impact cannot be scored: available citation signals are too thin — h-index 1 and i10-index 0 — to characterize influence. This axis would measure how widely and durably other researchers cite and build on this work.
Focus100/100
Focus scores 100 because every reviewed work sits in one declared field — machine learning spanning multimodal large language models and reinforcement learning — giving an undiluted research identity. The two 2026 catalog papers share one technical thread, GRPO-based policy optimization, applied to hallucination in multimodal LLMs (DEEPO) and credit assignment in LLM agents (TASPO). Wider topic counts, such as reinforcement learning in robotics (6) and topic modeling (6), hint at adjacent interests without diluting this core.
Collaboration51/100
Collaboration scores 51 from a mid-range network: 12 distinct co-authors appear in the recent record, and the frequent-coauthor list repeats Jingxiao Yang, Xuhong Zhang, and Jintao Chen, pointing to a small set of steady partners inside a larger group. On both 2026 catalog papers Gan is a middle author and not the corresponding author, consistent with contributing specialist expertise within team-led projects rather than leading the authorship.
Activity70/100
Activity scores 70 on current momentum: the snapshot records 10 works in the last three years, and countsByYear places all 10 in 2026, with both catalog papers appearing as 2026 arXiv preprints in quick succession. This is a concentrated recent burst rather than a long accumulation curve — the record is young — so the score reflects current tempo, not yet a sustained multi-year cadence.
Reproducibility55/100
Reproducibility scores 55, taken as a mean of 55 across one repro-scored paper. Both catalog papers carry a medium repro-readiness rating, and the reviews flag concrete gaps: DEEPO ships no code release alongside cross-table number inconsistencies, while TASPO reports strong ablations but artifacts that read as drafts. Methods are described in enough detail to attempt, yet independent replication would need extra effort.
Research summary
Wangjie Gan is a machine-learning researcher at Zhejiang University of Science and Technology whose recent work concentrates on reinforcement learning for large language models; OpenAlex field data records a single field spanning multimodal large language models and reinforcement learning. The 2026 catalog holds two arXiv preprints, with Gan as a middle author on each. DEEPO proposes dual-entropy enhanced policy optimization, combining semantic and Rényi entropy to curb hallucination in multimodal LLMs under GRPO-style training; Evolve finds the evidence unusually rigorous but flags cross-table number inconsistencies and no code release. TASPO reconciles process supervision with outcome-based credit in agentic policy optimization, reporting a +10.6% average gain across nine agent-RL settings with strong ablations, though draft artifacts lower trust. Broader topic signals — topic modeling (6), reinforcement learning in robotics (6), domain adaptation and few-shot learning (5), multimodal applications (4), and NLP techniques (3) — suggest an applied machine-learning span around this RL-for-LLM core. Output is recent and concentrated: 10 works in the last three years, all dated 2026. Bibliometric standing is early-stage (h-index 1, i10-index 0), reproducibility is rated medium, and the collaboration network covers 12 distinct co-authors.
Collaboration pattern
Wangjie Gan works within a moderately sized network: 12 distinct co-authors appear in the recent record, and the frequent-coauthor list repeats Jingxiao Yang, Xuhong Zhang, and Jintao Chen, indicating a few steady partners inside a larger team. On both 2026 catalog papers Gan is a middle author and not the corresponding author — a pattern consistent with contributing specialized reinforcement-learning expertise within team-based projects; other recurring names include Wenqi Zhang, Yingxuan Zhuang, and Fan Wang.
Research keywords
Highlights
Dual-entropy policy optimization against MLLM hallucination
Reconciling process supervision with outcome credit in LLM agents
Timeline
- 2026Posted TASPO to arXiv: process supervision reconciled with outcome-based credit in agentic policy optimization, reporting a +10.6% average gain over nine agent-RL settings.
- 2026Posted DEEPO to arXiv: dual-entropy enhanced policy optimization combining semantic and Rényi entropy to mitigate hallucination in multimodal LLMs.
- 2026OpenAlex snapshot logged 10 works dated 2026 alongside 12 distinct co-authors — the most active and collaborative year visible for this profile.
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 derived from OpenAlex only; no ORCID is linked, so this profile is name-resolved from public bibliographic data and same-name authors may be merged or split.
- Citation-impact data is insufficient (h-index 1, i10-index 0), so no impact score is reported.
- Only the two papers reviewed by Evolve have been quality-assessed; the remaining works behind the topic counts are unaudited.
- Both catalog papers are arXiv preprints, not yet confirmed in peer-reviewed venues.