Zhenbin Wang

Sichuan University · China

Unclaimed · AI-structured profileOpenAlex

Sichuan University researcher driving training-free efficiency and robustness in vision-language models

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-15) 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 scoring is unavailable: the current OpenAlex snapshot lacks complete venue and publication-stage data for a computed score, though three 2026 works are listed. This axis would measure productivity — the number, type, and venue quality of works over time.

Impact—

Impact cannot be scored yet: recorded h-index and i10-index are both 0 and citation counts are not available in the current data. This axis would measure citations, uptake, and the scholarly influence of the works.

Focus0/100

The focus score of 0 mirrors an even field distribution: efficient VLM inference, test-time adaptation/robustness, and vision-language models each hold about one third (0.333). Topic counts spread similarly across Advanced Neural Network Applications, Visual Attention and Saliency Detection, and Multimodal Machine Learning Applications (2 each). The methodological line is coherent, but coverage is dispersed across subfields.

Collaboration51/100

Collaboration scores 51, drawn from 12 distinct coauthors to date. The recurring cluster — Lei Zhang, Lituan Wang, Wei Huang, Yan Wang, Zhenwei Zhang, and Zhao Zhang — anchors the record, and all three cataloged 2026 preprints are first-author, non-corresponding works, consistent with leading experiments and writing within a stable team.

Activity40/100

Activity scores 40 on a thin base: countsByYear records 3 works in 2026 and recent3yWorks is 3, with no earlier years in the snapshot. All three outputs landed in a single year, so the cadence looks productive, but the record is too short to demonstrate sustained multi-year output.

Reproducibility70/100

Reproducibility is the strongest axis at 70, computed across 3 papers (reproMean 70). StackTok is rated high readiness — training-free with a well-specified submodular facility-location objective — while MASA and AREA are medium, with Evolve noting thin margins, modest gains, and missing statistical testing or latency numbers.

Research summary

Zhenbin Wang is a researcher at Sichuan University whose cataloged work centers on making vision-language and multimodal LLM inference more efficient and more robust — notably without retraining. Three first-author arXiv preprints, all dated 2026, define the current record. StackTok introduces a budget-adaptive visual token selector built on submodular facility location, balancing query relevance and visual coverage; Evolve rates it the top training-free baseline and marks its reproducibility readiness as high. MASA targets wild test-time adaptation by adding sparse MLLM semantic anchors and prototype memory to break the self-referential adaptation loop, with consistent but modest ImageNet-C gains. AREA is a training-free evidence-highlighting intervention that adaptively gates, scales, and refreshes attention, beating LoT across all nine KB-VQA settings. The signature across all three is inference-time, training-free intervention — token selection, semantic anchoring, evidence gating — applied from different angles. The OpenAlex field distribution splits evenly (about 0.333 each) across efficient VLM inference, test-time adaptation/robustness, and vision-language models, and the focus score of 0 reflects that breadth. Recorded h-index and i10-index are both 0, with 12 distinct coauthors, so this is an early-stage, first-author-driven profile built on a short but methodologically coherent record.

Collaboration pattern

OpenAlex records 12 distinct coauthors, and the collaboration score is 51. The recurring cluster — Lei Zhang, Lituan Wang, Wei Huang, Yan Wang, Zhenwei Zhang, and Zhao Zhang — anchors the record, and all three cataloged preprints are first-author, non-corresponding works, consistent with leading experiments and writing within a stable team rather than solo or corresponding-author roles. Co-author roles beyond these names are not documented in the current data.

Research keywords

visual token pruningvision-language modelstraining-free inference accelerationtest-time adaptationdistribution shiftmultimodal large language modelsknowledge-based VQAattention sinksretrieval-augmented generationsubmodular facility locationprototype memorysemantic anchoring

Highlights

Budget-adaptive visual token selection for faster VLM inference
First-author 2026 arXiv preprint proposing a training-free selector grounded in submodular facility location that trades off query relevance and visual coverage. Evolve's verdict: it tops all tested training-free baselines and is principled and well specified, though margins are thin and no latency is reported. Reproducibility readiness: high.
Sparse MLLM anchors break the self-referential loop in wild test-time adaptation
First-author 2026 arXiv preprint (MASA) adding sparse MLLM semantic anchors and a prototype memory to wild test-time adaptation so the model stops referencing itself under distribution shift. Evolve finds ImageNet-C gains consistent but modest, and notes the lack of statistical testing. Reproducibility readiness: medium.
Training-free evidence selection for multimodal LLMs across layers and attention sinks
First-author 2026 arXiv preprint (AREA): an inference-time intervention that adaptively gates, scales, and refreshes evidence highlighting for knowledge-based VQA. Per Evolve, it beats LoT on all nine KB-VQA settings, with modest, untested gains. Reproducibility readiness: medium.

Timeline

2026StackTok, a first-author arXiv preprint on budget-adaptive visual token selection for VLM inference acceleration, enters the record with high reproducibility readiness.
2026MASA, a first-author arXiv preprint adding sparse MLLM semantic anchors to wild test-time adaptation, is cataloged with medium reproducibility readiness.
2026AREA, a first-author arXiv preprint on adaptive evidence selection for multimodal LLMs, is cataloged with medium reproducibility readiness.
2026All three cataloged works fall in 2026, giving recent3yWorks = 3 and a countsByYear entry of 3 for 2026 only.

Papers in the Evolve catalog

2026-09-17first author
StackTok: Accelerating VLMs Inference with Budget-Adaptive Visual Token Selection
2026-09-15first author
Sparse MLLM Anchors, Dense Adaptation: Breaking the Self-Referential Loop in Wild Test-Time Adaptation
2026-09-15first author
Layers, Sinks, and Scaling: Adaptive Evidence Selection for Multimodal Large Language Models

Frequent co-authors

Lei Zhang×3 Lituan Wang×3 Wei Huang×3 Yan Wang×3 Zhenwei Zhang×1 Zhao Zhang×1 Zhenbin Wang×1 Lei Zhang×1 Lituan Wang×1 Wei Huang×1 Yan Wang×1 Zhenwei Zhang×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 dimensions are computed from OpenAlex only; no ORCID is linked, so the profile is name-resolved from public bibliographic records and may merge or split same-name authors.
Impact is unscored and recorded h-index and i10-index are both 0; citation coverage for these 2026 arXiv preprints is not yet available.
All three cataloged papers are 2026 arXiv preprints; peer-review status, venues, latency numbers, and testing details come from Evolve review notes only.
Only three works are in scope, so the topic mix, the even three-way field split, and the focus score of 0 reflect a small sample and will shift as the record grows.