Peishun Liu

—

Unclaimed · AI-structured profileOpenAlex

Multimodal vision-language model researcher: training-free decoding-time fixes for LVLM hallucination and gender bias

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

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—

Not yet scored. Beyond the 3 works OpenAlex records for 2026, no reliable volume signal exists (h-index 0, i10-index 0). This axis would measure total works, venue mix and authorship position over time.

Impact—

Not yet scored. Citation data is absent — h-index and i10-index both read 0, unsurprising for 2026 arXiv-first work. This axis would track citations, influential citations and uptake of the cataloged papers as the record matures.

Focus100/100

Scores 100/100. OpenAlex topic tags are led by Multimodal Machine Learning Applications (3 works), and the only field assignment in fieldDist is machine learning (multimodal vision-language models). Both cataloged papers target large vision-language models — one diagnoses hallucinations via contrastive decoding probes, the other mitigates gender bias — a tightly concentrated early agenda on LVLM reliability and fairness.

Collaboration44/100

Scores 44/100, with 8 distinct coauthors recorded. The recurring names are Zhipeng Zhao, Ruichun Tang, Wenxu Wang, Zhaoqiang Wei and Youwei Zhao, pointing to a small, stable team. Both 2026 catalog papers list this researcher as a middle author and not corresponding, consistent with contributing within a team rather than leading an independent group so far.

Activity40/100

Scores 40/100. OpenAlex records 3 works in 2026 (recent3yWorks = 3), and countsByYear contains that single year, so the publication record spans just one active year. Both catalog papers are 2026 arXiv postings. The score reflects a recent start rather than a sustained multi-year cadence, and should be re-scored as later years accrue.

Reproducibility55/100

Scores 55/100, the mean repro-readiness across 1 assessed paper (reproPapers = 1). The contrastive-decoding hallucination study is rated medium: Evolve's review calls it candid but flags that the 2x MMHalBench gain looks judge-confounded and probe AUROC is modest. With only one assessment, this axis remains provisional until more outputs are evaluated.

Research summary

Peishun Liu is an early-career machine learning researcher whose visible work concentrates on large vision-language models (LVLMs); this profile is name-resolved from OpenAlex bibliographic data plus two 2026 arXiv papers in the Evolve catalog. The record shows a sharply focused agenda: OpenAlex tags Multimodal Machine Learning Applications as the dominant topic (3 works) and the only field label is machine learning (multimodal vision-language models), yielding a 100/100 focus score. The two catalog papers attack LVLM trustworthiness from complementary angles — one probes the visual grounding failures behind hallucinations via contrastive decoding (SAFE), a training-free decoding-time intervention; the other, ViD, mitigates gender bias with a vision-dominant strategy. Both list Liu in a middle author position, neither as corresponding author. Activity is recent: 3 works recorded in 2026, the only publication year in countsByYear, with 8 distinct coauthors including recurring names Zhipeng Zhao, Ruichun Tang, Wenxu Wang, Zhaoqiang Wei and Youwei Zhao. Citation impact cannot yet be assessed — h-index and i10-index both read 0 — and the single reproducibility assessment (medium readiness) averages to 55/100; Evolve's review notes the reported 2x MMHalBench gain may be judge-confounded and probe AUROC is modest. Expect this early-stage profile to evolve as indexing catches up.

Collaboration pattern

OpenAlex records 8 distinct coauthors. The recurring names are Zhipeng Zhao (listed twice), Ruichun Tang (twice), Wenxu Wang, Zhaoqiang Wei and Youwei Zhao — a compact, apparently stable group. Peishun Liu's own name also appears in the coauthor list, a likely OpenAlex name-resolution artifact worth noting. Both 2026 catalog papers place Liu in a middle author position, not corresponding, suggesting contribution within a team rather than independent leadership at this stage.

Research keywords

LVLM hallucinationcontrastive decodingvisual groundingtoken-level diagnosisdecoding-time interventiontraining-free mitigationlarge vision-language modelsmultimodal reasoninggender bias mitigationvision-dominant bias mitigationmultimodal machine learningAI ethics and fairness

Highlights

Contrastive decoding probes trace LVLM hallucinations to visual grounding failures
This 2026 arXiv paper (middle authorship) introduces SAFE, which contrasts vision-on vs. vision-off token log-probabilities as a token-level, training-free decoding-time probe and mitigation for hallucination in large vision-language models. Evolve's verdict is cautiously positive — a candid paper — but flags that the 2x MMHalBench gain looks judge-confounded and the probe's AUROC is modest; reproducibility is rated medium.
ViD: a vision-dominant strategy for mitigating gender bias in LVLMs
The second 2026 arXiv paper in the catalog, 'ViD: Vision-Dominant Gender Bias Mitigation for Large Vision-Language Models', extends the researcher's focus from hallucination to fairness, again with middle authorship. Evolve has not yet published a verdict, keywords or a reproducibility rating for this paper.
A coherent, training-free agenda for LVLM trustworthiness
Taken together, the 2026 output — 3 works recorded in OpenAlex, with Multimodal Machine Learning Applications as the top topic — forms one coherent line: decoding-time, training-free interventions for large vision-language model reliability (hallucination) and fairness (gender bias), reflected in the perfect 100/100 focus score.

Timeline

2026Records 3 works in OpenAlex for 2026 — the only publication year in the current snapshot — with 8 distinct coauthors and Multimodal Machine Learning Applications as the leading topic.
2026Posts 'What Do Hallucinations Reveal About Multimodal Reasoning?' to arXiv, diagnosing visual grounding failures via contrastive decoding probes (SAFE), in a middle author position.
2026Co-authors 'ViD: Vision-Dominant Gender Bias Mitigation for Large Vision-Language Models' on arXiv, extending the agenda from hallucination to gender bias, again as a middle author.

Papers in the Evolve catalog

Frequent co-authors

Zhipeng Zhao×2 Wenxu Wang×1 Ruichun Tang×1 Zhipeng Zhao×1 Zhaoqiang Wei×1 Peishun Liu×1 Youwei Zhao×1 Ruichun Tang×1

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 come from OpenAlex only (identity: openalex); h-index and i10-index are 0 and no citation counts are available, so output and impact are left unscored.
Institution, country and ORCID are not on record; 'Peishun Liu' is name-resolved from public bibliographic data and may be a common name.
Only 2 papers from the Evolve catalog inform this profile (both 2026, arXiv); one lacks a review verdict, keywords and reproducibility rating.
The reproducibility score averages a single assessed paper, so the 55/100 figure is provisional.