Early-career LVLM researcher: first-author work on hallucination diagnosis and decoding-time bias mitigation
2
papers here
1
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-20) 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 yet: this axis would aggregate total works and per-year production counts, but the snapshot carries no reliable cumulative work total for this identity beyond the 2026 cluster, so the axis stays unscored.
Impact—
Impact is not scored: this axis measures citations, h-index, and i10-index. The snapshot records 0 for both indices, which for a 2026-only profile reflects missing citation coverage rather than measured influence, so no estimate is offered.
Focus100/100
Focus is the strongest axis (100/100). The OpenAlex field distribution lists a single field — machine learning with multimodal vision-language models — and 3 of the profile's topic mentions fall under Multimodal Machine Learning Applications, with single mentions each for topic modeling, neural network applications, child/animal learning development, and AI ethics. Both cataloged 2026 papers target large vision-language models, indicating a tightly concentrated research agenda.
Collaboration44/100
Collaboration scores 44/100. The OpenAlex window records 8 distinct co-authors, and the frequent list is anchored by Peishun Liu and Ruichun Tang (each recurring), alongside Wenxu Wang, Zhaoqiang Wei, and Youwei Zhao. Both cataloged papers are first-authored but not corresponding, consistent with a contributing member of a small, stable team rather than a solo or PI-style working mode.
Activity40/100
Activity scores 40/100. The countsByYear series contains a single year — 3 works in 2026 — and recent3yWorks is 3, so output is recent but thin. All cataloged production is first-author arXiv preprints from 2026; there is no visible publication history before 2026 in this snapshot, which keeps the activity estimate conservative and early-stage.
Reproducibility55/100
Reproducibility scores 55/100, the mean over 1 assessed paper. Evolve rates the hallucination-probe paper's reproducibility readiness as medium; the verdict praises a candid contrastive decoding analysis while cautioning that the reported MMHalBench gain may be judge-confounded and that probe AUROC is modest. With only one paper assessed, this score is indicative rather than stable.
Research summary
Zhipeng Zhao is an early-stage machine learning researcher working on large vision-language models (LVLMs), name-resolved from OpenAlex public bibliographic data; no institution, country, or ORCID is attached to the profile. The visible record is concentrated in 2026, with 3 works recorded that year and both cataloged papers as first-author arXiv preprints. The first, 'What Do Hallucinations Reveal About Multimodal Reasoning?', proposes a contrastive decoding probe that compares token log-probabilities with vision on versus off to localize visual grounding failures, and applies a training-free, decoding-time mitigation for LVLM hallucination; Evolve's review calls the paper candid but flags that the reported 2x MMHalBench gain may be judge-confounded and that probe AUROC is modest, with medium reproducibility readiness. The second, 'ViD: Vision-Dominant Gender Bias Mitigation for Large Vision-Language Models', extends this agenda to fairness, targeting gender bias mitigation in LVLMs; it has not yet been fully reviewed by Evolve. Topic signals are consistent: Multimodal Machine Learning Applications dominates, with side mentions of topic modeling, neural network applications, learning development, and AI ethics. Axes show perfect focus (100) but modest activity (40) and collaboration (44), reflecting a short record with 8 distinct co-authors; output and impact are unscored because citation and cumulative-work data are not yet available. This is a full but very young profile whose trajectory is defined by decoding-time interventions for trustworthy multimodal models.
Collaboration pattern
OpenAlex records 8 distinct co-authors in the current window. The frequent-coauthor list is anchored by Peishun Liu and Ruichun Tang (the only names appearing more than once), with Wenxu Wang, Zhaoqiang Wei, and Youwei Zhao also present; a self-entry suggests minor disambiguation noise. Both cataloged papers are first-authored without corresponding authorship, consistent with a hands-on contributor inside a small, stable team rather than a solo or supervising role.
Research keywords
Highlights
Contrastive decoding probes trace LVLM hallucinations to visual grounding
ViD: vision-dominant gender bias mitigation for large vision-language models
Second first-author 2026 arXiv preprint proposes ViD, a vision-dominant approach to mitigating gender bias in large vision-language models, extending the same trustworthy-multimodality agenda from hallucination to fairness. It is not yet fully reviewed by Evolve, so verdict, keywords, and reproducibility readiness are pending.
Timeline
- 2026First cataloged first-author arXiv preprint: contrastive decoding probes diagnosing the visual grounding failures behind LVLM hallucinations.
- 2026Second first-author preprint, ViD, extends the agenda to vision-dominant gender bias mitigation for large vision-language models.
- 2026OpenAlex snapshot logs 3 works in 2026 with 8 distinct co-authors; h-index and i10-index remain 0, marking a debut-year 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.
- Profile is name-resolved from OpenAlex public bibliographic data; institution, country, and ORCID are unavailable.
- Bibliometric scores come from the OpenAlex snapshot only; h-index and i10-index read 0, and output and impact remain unscored pending citation coverage.
- Only one of the two cataloged papers has complete Evolve review metadata (verdict, keywords, reproducibility readiness); the ViD preprint is pending.
- The frequent-coauthor list contains duplicates and a self-entry, indicating author-disambiguation noise in the underlying records.