Yan Wang

Sichuan University · China

Unclaimed · AI-structured profileOpenAlex

Yan Wang (Sichuan University): training-free efficiency and robustness for vision-language and multimodal 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—

Not yet scored: this OpenAlex-derived snapshot records only 3 recent works and no verified total publication count, so cumulative output and its trajectory cannot be measured. This axis would track publication volume, venue mix, and productivity trends over time.

Impact—

Not yet scored: no citation counts or venue-level indicators exist for these 2026 arXiv preprints, and h-index and i10-index are recorded only as zeros. This axis would measure citations, uptake, and influence on subsequent research.

Focus0/100

The focus score is 0 because the field distribution is perfectly even: efficient VLM inference, test-time adaptation/robustness, and vision-language modeling each hold 0.333 of the record. Spread this wide lowers concentration, yet the strands are methodologically coherent — all three 2026 papers apply training-free, inference-time interventions to multimodal models, from token selection to semantic anchoring to evidence gating.

Collaboration51/100

Collaboration scores 51 on a 0-100 scale, based on 12 distinct coauthors in the record. Recurring names include Zhenbin Wang, Lei Zhang, Lituan Wang, Wei Huang, Zhenwei Zhang, and Zhao Zhang, suggesting a stable team around Sichuan University. On all three 2026 preprints Yan Wang is a middle author and not the corresponding author, indicating a contributor role rather than project leadership so far.

Activity40/100

Activity scores 40: the snapshot lists 3 works in the last three years, with countsByYear showing all 3 in 2026. That is a concentrated recent burst rather than sustained volume, consistent with an early-stage record, and h-index and i10-index both stand at 0, reflecting how new the visible output is.

Reproducibility70/100

Reproducibility averages 70 across the 3 papers Evolve assessed. StackTok is rated high readiness — training-free, principled, and well specified — while MASA and AREA are rated medium, since their gains are consistent but modest and lack statistical testing. The training-free design choices across the portfolio support reimplementation.

Research summary

Yan Wang is a researcher at Sichuan University whose recent work centers on making vision-language and multimodal large language models more efficient, adaptive, and robust at inference time. The Evolve catalog holds three 2026 arXiv preprints with Yan Wang as a middle author: StackTok (budget-adaptive visual token selection for training-free VLM acceleration, rated high reproducibility readiness), MASA (sparse MLLM semantic anchors that break the self-referential loop in wild test-time adaptation), and AREA (adaptive evidence selection that beats LoT on all nine knowledge-based VQA settings). Topic records — advanced neural network applications, visual attention and saliency detection, and multimodal machine learning applications (2 counts each), plus topic modeling and domain adaptation/few-shot learning (1 each) — match this orientation. The field distribution splits evenly (0.333 each across efficient VLM inference, test-time adaptation/robustness, and vision-language models), producing a concentration score of 0, though all three strands share one methodology: training-free, inference-time intervention. Collaboration is moderate (score 51; 12 distinct coauthors), with recurring partners including Zhenbin Wang, Lei Zhang, Lituan Wang, Wei Huang, Zhenwei Zhang, and Zhao Zhang. Activity is early-stage: 3 works in the past three years, all dated 2026, with h-index and i10-index recorded as 0. This is an early-stage profile built solely from OpenAlex bibliographic data; the name is common, so resolution is bibliographic rather than biographical.

Collaboration pattern

The record suggests a compact, stable collaboration cluster: 12 distinct coauthors, with Zhenbin Wang, Lei Zhang, Lituan Wang, Wei Huang, Zhenwei Zhang, and Zhao Zhang recurring across the reviewed work. Yan Wang appears as a middle author on all three 2026 arXiv preprints and as corresponding author on none, consistent with a contributing member of a Sichuan University team rather than its lead.

Research keywords

Visual token pruningVision-language modelsTraining-free inference accelerationTest-time adaptationDistribution shiftMultimodal large language modelsKnowledge-based VQAAttention sinksRetrieval-augmented generationPrototype memory

Highlights

Training-free token selection tops VLM acceleration baselines
StackTok performs budget-adaptive visual token selection for training-free VLM inference acceleration, drawing on submodular facility location, query relevance, and visual coverage. Evolve finds it tops all tested training-free baselines and calls it principled and well specified, but notes thin margins and no latency reporting; reproducibility readiness is high.
Sparse MLLM anchors break the self-referential loop in wild TTA
MASA adds sparse MLLM semantic anchors and prototype memory to wild test-time adaptation, breaking the self-referential loop under distribution shift. Evolve reports consistent but modest gains on ImageNet-C and flags the absence of statistical testing; reproducibility readiness is medium.
Adaptive evidence selection wins all nine KB-VQA settings
AREA is a training-free inference-time intervention for multimodal large language models that adaptively gates, scales, and refreshes evidence highlighting around attention sinks, with retrieval-augmented context for knowledge-based VQA. Evolve reports it beats LoT on all nine settings, with modest, untested gains; reproducibility readiness is medium.

Timeline

2026StackTok appears on arXiv, introducing budget-adaptive visual token selection to accelerate VLM inference; Evolve later rates its reproducibility readiness as high.
2026MASA appears on arXiv, adding sparse MLLM semantic anchors to wild test-time adaptation and reporting consistent but modest gains on ImageNet-C.
2026AREA appears on arXiv, delivering training-free adaptive evidence selection that beats LoT across all nine knowledge-based VQA settings.
2026All three catalog papers land in a single year, forming the entire visible recent output (3 works in the last three years) for this early-stage record.

Papers in the Evolve catalog

Frequent co-authors

Zhenbin Wang×3 Lei Zhang×3 Lituan Wang×3 Wei Huang×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.

Bibliometrics come from OpenAlex only; h-index and i10-index are recorded as 0 with no citation data yet, so output and impact are unscored.
"Yan Wang" is a common name; this profile is name-resolved from public bibliographic data and may mix records of different researchers with the same name.
All catalog papers are 2026 arXiv preprints with middle authorship; no peer-reviewed venue, award, or career history information is in scope.
The reproducibility score of 70 reflects Evolve's assessment of only 3 papers and may change as the catalog grows.