Lei Zhang

Sichuan University · China

Unclaimed · AI-structured profileOpenAlex

Sichuan University researcher advancing training-free efficiency for vision-language models and test-time robustness.

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 is not yet scored: this name-resolved OpenAlex profile lacks reliable publication counts (h- and i10-index unpopulated), so this axis — measuring cumulative papers and yearly output — cannot be computed without confirmed record linkage.

Impact—

Impact is not yet scored: citation counts are unavailable for this name-resolved profile, so this axis — measuring citations, h-index trajectory, and influence on follow-up work — awaits verified attribution of the 2026 arXiv papers.

Focus0/100

The focus score is 0, reflecting a deliberately broad portfolio: the field distribution splits evenly at 0.33 across efficient VLM inference, test-time adaptation and robustness, and vision-language modelling, with no dominant field. OpenAlex topics reinforce this spread — topic modelling, neural network applications, visual saliency, and multimodal learning each appear twice, alongside one particle-physics topic — suggesting early-stage exploration rather than a settled specialisation.

Collaboration51/100

Collaboration scores 51, based on 12 distinct coauthors in the snapshot. The reviewed 2026 papers list Lei Zhang in a middle author position on all three, none as corresponding author, consistent with a contributor role inside a compact Sichuan University group. Recurring partners include Zhenbin Wang, Lituan Wang, Wei Huang, Yan Wang, Zhenwei Zhang, and Zhao Zhang, indicating a small, stable core team rather than a wide external network.

Activity52/100

Activity scores 52 from a thin recent window: OpenAlex records 5 works in the last three years, all logged in 2026 (countsByYear shows 2026: 5). The Evolve-reviewed sample — three arXiv preprints on VLM efficiency and test-time adaptation — confirms steady 2026 output, but the absence of earlier years in the snapshot means cadence before 2026 cannot be characterised.

Reproducibility70/100

Reproducibility scores 70, the mean readiness across 3 reviewed papers. StackTok rates high — training-free, principled, and well specified — while MASA and AREA rate medium: gains are consistent but modest, and Evolve notes missing statistical testing and, for StackTok, no reported latency. Overall the work favours training-free, specified methods that are straightforward to re-implement.

Research summary

Lei Zhang is a researcher at Sichuan University whose visible work centres on making vision-language and multimodal large language models faster and more robust at inference time. This is a name-resolved profile built from OpenAlex bibliographic data; 'Lei Zhang' is a common name, so record linkage should be treated as provisional. The Evolve-reviewed sample comprises three 2026 arXiv preprints, all with Zhang in a middle author position: StackTok (training-free, budget-adaptive visual token selection for VLM inference), MASA (sparse multimodal-LLM semantic anchors for wild test-time adaptation), and AREA (adaptive evidence highlighting across layers, attention sinks, and scaling for knowledge-based VQA). OpenAlex attributes 5 works to this profile in the last three years, all in 2026, with 12 distinct coauthors and a field distribution split evenly (0.33 each) across efficient VLM inference, test-time adaptation and robustness, and vision-language modelling — a deliberately broad, early-stage portfolio, consistent with a focus score of 0. Reproducibility is the strongest measured axis (70/100, the mean across 3 reviewed papers), reflecting a training-free, well-specified methodological style; StackTok rates high while MASA and AREA rate medium. Output and impact are unscored because citation and publication-count data are unavailable for this profile, and no peer-reviewed venue record appears in the input data.

Collaboration pattern

Bibliographic data shows a compact, stable collaboration pattern: 12 distinct coauthors, with repeated pairing alongside Zhenbin Wang, Lituan Wang, Wei Huang, Yan Wang, Zhenwei Zhang, and Zhao Zhang. On all three Evolve-reviewed 2026 arXiv papers, Lei Zhang appears in a middle author position and none as corresponding author, suggesting a hands-on contributor within a Sichuan University team rather than a corresponding-author lead. No international or cross-institution patterns can be confirmed from the snapshot.

Research keywords

visual token pruningvision-language modelstraining-free inference accelerationsubmodular facility locationtest-time adaptationdistribution shiftprototype memorymultimodal large language modelsknowledge-based VQAattention sinksevidence highlightingretrieval-augmented generation

Highlights

StackTok: budget-adaptive visual token selection for faster VLM inference
A training-free visual token selector that tops all tested training-free baselines; Evolve calls it principled and well specified (reproducibility: high), while noting thin margins and no reported latency.
MASA breaks the self-referential loop in wild test-time adaptation
Adds sparse multimodal-LLM semantic anchors to wild TTA via prototype memory, delivering consistent but modest ImageNet-C gains; Evolve flags the absence of statistical testing (reproducibility: medium).
AREA: adaptive evidence selection across layers, sinks, and scaling
Training-free gating, scaling, and refreshing of evidence highlighting for multimodal LLMs; it beats LoT on all nine KB-VQA settings, though Evolve notes the gains are modest and untested (reproducibility: medium).

Timeline

2026StackTok introduces budget-adaptive visual token selection for training-free VLM inference acceleration (arXiv preprint).
2026Sparse MLLM Anchors, Dense Adaptation (MASA) adds sparse multimodal-LLM semantic anchors to wild test-time adaptation (arXiv preprint).
2026AREA presents adaptive evidence selection for multimodal LLMs across layers, attention sinks, and scaling (arXiv preprint).
2026OpenAlex attributes 5 works to this profile in 2026, including the three Evolve-reviewed preprints.

Papers in the Evolve catalog

Frequent co-authors

Zhenbin Wang×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.

Name-resolved profile: 'Lei Zhang' is a common name; identity is matched from OpenAlex bibliographic data only, and no ORCID is provided.
Bibliometrics (coauthor counts, topics, activity) come from OpenAlex only; citation, h-index, and publication-volume data are unavailable, so output and impact axes are unscored.
Researcher-level evidence currently covers three 2026 arXiv preprints reviewed by Evolve; no peer-reviewed venue record appears in the input data.
The frequent-coauthor list contains repeated entries (including one self-entry), so collaborator roles and seniority cannot be verified.