Wei Zhang

—

Unclaimed · AI-structured profileOpenAlex

Researcher in AI safety evaluation and multimodal hallucination mitigation, studying LLM behavior under repeated abuse

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: publication volume, venue distribution, and yearly output counts are insufficient in the current data. This axis would measure total research productivity and how it trends over the researcher's career.

Impact—

Not yet scored: citation counts and other influence measures are unavailable in the current snapshot. This axis would measure how far the two catalogued works propagate through downstream research and practice.

Focus100/100

The focus dimension scores 100/100, the maximum. The underlying field distribution contains a single classified field, 'nlp / ai safety evaluation', drawn from one classified paper. This indicates complete concentration in that field among classified output, though the base is thin: the second catalogued paper (SAVOR) carries no field label, and its reinforcement-learning and visual-grounding topics could broaden the picture once classified. Treat the perfect score as evidence of narrow classified coverage, not long-term specialization.

Collaboration51/100

The collaboration dimension scores 51/100, near the scale midpoint, from 12 distinct co-authors in the record. The list includes William Guey, Pierrick Bougault, Yi Wang, Agoston Bodo, and Jose O Gomes among others, with no single partner dominating. Across the two catalogued papers Zhang is middle author on one and last author on the other, and corresponding author on neither — a contributing rather than leading position.

Activity—

Not yet scored: recent publication cadence and submission patterns cannot be computed from the data. This axis would measure ongoing research tempo, such as outputs per year and gaps between publications.

Reproducibility100/100

The reproducibility dimension scores 100/100, computed from a mean reproducibility of 100 across exactly one reviewed paper (reproPapers = 1). That paper, 'How AI Assistants Respond to Repeated Abuse', carries a high reproducibility-readiness rating, and Evolve's verdict calls its measurement careful and transparent. With a single-paper base, this score is an encouraging signal about methodological transparency, not a stable track record.

Research summary

Wei Zhang is an early-stage researcher working at the intersection of AI safety evaluation and multimodal learning. This profile is name-resolved from public bibliographic data; 'Wei Zhang' is a common name, so coverage should be read with that limitation in mind. The current evidence base consists of two 2026 arXiv papers. The first, 'How AI Assistants Respond to Repeated Abuse', places Zhang in a middle author position and examines how AI assistant configurations behave under repeated verbal abuse; Evolve's review concludes that refusal is not a single axis, that different assistant configurations diverge sharply under repeated abuse, and that the measurement is careful and transparent while the study remains exploratory and deployment-bound. The second, 'SAVOR: Self-Aware Visual Grounding via Confidence-Calibrated Reinforcement Learning for Multimodal Hallucination Mitigation', lists Zhang as last author and targets hallucination in vision-language grounding through confidence-calibrated reinforcement learning. Bibliometrically, the profile shows a maximal focus score (100) within NLP / AI safety evaluation, though only one paper carries a field classification; a collaboration score of 51 across 12 distinct co-authors; and a reproducibility score of 100 computed from a single reviewed paper. Output, impact, and activity axes lack sufficient data and are not scored.

Collaboration pattern

Zhang's recorded network contains 12 distinct co-authors with no single dominant partner: William Guey, Pierrick Bougault, Yi Wang, Agoston Bodo, Jose O Gomes, Zixiu Ding, Zilin Zhao, Yingjie He, Xinlang Kang, and Guansu Wang appear alongside Vitor D de Moura. The list mixes Western and Chinese given names, though no affiliation data is present. On both catalogued papers Zhang is a non-corresponding author — middle on the abuse study and last on SAVOR — consistent with a contributor or senior-adjacent role.

Research keywords

conversational AI safetyrefusal and disengagementLLM-as-a-judgemulti-turn dialogue evaluationbilingual benchmarkverbal abusemultimodal hallucination mitigationvisual groundingconfidence calibrationreinforcement learning

Highlights

Refusal is not one axis: assistant configs diverge under repeated abuse
In the arXiv study 'How AI Assistants Respond to Repeated Abuse' (2026), Zhang contributes in a middle author position to a bilingual, multi-turn dialogue evaluation of AI assistants under repeated verbal abuse, using LLM-as-a-judge measurement. Evolve's verdict: refusal is not a single axis — assistant configurations differ sharply under repeated abuse; measurement is careful and transparent, but the study is exploratory and deployment-bound. Reproducibility readiness is rated high.
Confidence-calibrated RL against multimodal hallucination (SAVOR)
On 'SAVOR: Self-Aware Visual Grounding via Confidence-Calibrated Reinforcement Learning for Multimodal Hallucination Mitigation' (arXiv, 2026), Zhang is the last-listed author. The paper proposes self-aware visual grounding trained with confidence-calibrated reinforcement learning to mitigate multimodal hallucination. It has not yet been reviewed by Evolve and carries no field, keyword, or reproducibility metadata in the catalog.

Timeline

2026First catalogued paper: 'How AI Assistants Respond to Repeated Abuse' appears on arXiv, with Zhang in a middle author position and a high reproducibility-readiness rating.
2026'SAVOR: Self-Aware Visual Grounding via Confidence-Calibrated Reinforcement Learning for Multimodal Hallucination Mitigation' appears on arXiv with Zhang as last author.
2026Both catalogued outputs land in the same year, making 2026 the entire observed window of this profile's evidence base — an early-stage record.

Papers in the Evolve catalog

Frequent co-authors

William Guey×1 Pierrick Bougault×1 Yi Wang×1 Agoston Bodo×1 Vitor D de Moura×1 Jos\'e O Gomes×1 Yi Wang×1 Zixiu Ding×1 Zilin Zhao×1 Yingjie He×1 Xinlang Kang×1 Guansu Wang×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.

Bibliometric scores come from an OpenAlex-derived public bibliometric snapshot only; output, impact, and activity axes lack underlying data and are unscored.
'Wei Zhang' is a very common name; this profile is name-resolved from public bibliographic data and may merge or omit same-name records.
Evidence covers two 2026 arXiv preprints only — no peer-reviewed venues, affiliation, country, or ORCID were provided.
The reproducibility score of 100 rests on a single reviewed paper (reproPapers = 1) and should not be read as an established track record.