Wei Huang

Sichuan University · China

Unclaimed · AI-structured profileOpenAlex

Wei Huang (Sichuan University): training-free VLM inference acceleration and test-time adaptation

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 scorable: the underlying OpenAlex data lacks the publication-volume metrics this axis needs. It would measure cumulative output, venue quality, and productivity over time; only three 2026 arXiv preprints are currently on record.

Impact—

Impact is not yet scorable: h-index and i10-index are 0, and the three 2026 preprints are too recent to have accumulated citations. This axis would measure citation footprint and scholarly influence once citations accrue.

Focus0/100

The raw focus score is 0, the lowest axis. OpenAlex splits this profile evenly across three fields — efficient VLM inference, test-time adaptation and robustness, and vision-language models, each at 0.333 — and lists five topics with thin counts (three at 2, two at 1). A real thread exists — training-free, inference-time methods for multimodal models — but coverage is still too even to register as concentration.

Collaboration51/100

The raw collaboration score is 51, with 12 distinct coauthors on record. All three 2026 papers share one compact recurring team — Zhenbin Wang, Lei Zhang, Lituan Wang, Yan Wang, Zhenwei Zhang, and Zhao Zhang appear repeatedly across the coauthor lists — pointing to a stable, likely single-institution cluster at Sichuan University with little visible external networking so far.

Activity40/100

The raw activity score is 40. The counts-by-year record shows all measurable output in a single year — 3 works in 2026 — and OpenAlex lists 3 works over the last three years, matching the three arXiv preprints in the Evolve catalog. The pattern is a recent burst: momentum is visible, but no multi-year publication record yet exists to lift the score.

Reproducibility70/100

The raw reproducibility score is 70, averaged over 3 reviewed papers — the strongest axis on this profile. StackTok is rated high readiness: principled and well specified, though margins are thin and no latency is reported. MASA and AREA are rated medium: gains are consistent but modest and untested, with no statistical testing. Training-free designs and explicit baselines support replication.

Research summary

Wei Huang is a researcher at Sichuan University whose recent work makes vision-language and multimodal large language models faster, more robust, and more faithful at inference time. The Evolve catalog holds three 2026 arXiv preprints: StackTok, a training-free, budget-adaptive visual token selector for accelerating VLM inference; MASA, which adds sparse MLLM semantic anchors to wild test-time adaptation to break the self-referential loop; and AREA, a training-free adaptive evidence-selection method for knowledge-based VQA. OpenAlex records 3 works in the last three years, 12 distinct coauthors, and an h-index and i10-index of 0, so output and impact cannot yet be scored and this remains an early-stage profile. The bibliometric field distribution splits evenly, about one third each across efficient VLM inference, test-time adaptation and robustness, and vision-language models. Reproducibility is comparatively strong, averaging 70 across the three reviewed papers, with StackTok rated high readiness. Author positions are middle or last, none corresponding, fitting a contributing member of a small Sichuan University collaboration cluster.

Collaboration pattern

All three 2026 papers come from one compact team: Zhenbin Wang, Lei Zhang, Lituan Wang, Yan Wang, Zhenwei Zhang, and Zhao Zhang recur across the coauthor records, within a total of 12 distinct coauthors. This suggests a stable, likely single-institution cluster at Sichuan University with little visible external collaboration. Wei Huang appears in middle or last author positions, never as corresponding author — consistent with a contributing technical member rather than the cluster lead.

Research keywords

efficient VLM inferencevisual token pruningtraining-free inference accelerationtest-time adaptationdistribution shift robustnessmultimodal large language modelsknowledge-based visual question answeringattention sinksevidence highlightingretrieval-augmented generationsubmodular facility locationprototype memory

Highlights

StackTok: budget-adaptive visual token selection for training-free VLM acceleration
A training-free token selector built on submodular facility location, balancing query relevance and visual coverage under a token budget. Evolve's review finds it tops all tested training-free baselines with a principled, well-specified method — though margins are thin and no latency numbers are reported. Reproducibility readiness: high.
MASA: sparse MLLM anchors break the self-referential loop in wild test-time adaptation
The paper injects sparse multimodal-LLM semantic anchors into wild test-time adaptation, using a prototype memory to counter distribution shift. Evolve finds the gains on ImageNet-C consistent but modest, and notes the absence of statistical testing. Reproducibility readiness: medium. Wei Huang is the last-listed author.
AREA: adaptive evidence selection across layers, sinks, and scaling for MLLMs
A training-free, inference-time intervention that adaptively gates, scales, and refreshes evidence highlighting for knowledge-based VQA. Evolve reports it beats LoT on all nine KB-VQA settings, while cautioning that the gains are modest and untested; attention sinks inform the design. Reproducibility readiness: medium.

Timeline

2026StackTok preprint appears on arXiv: training-free, budget-adaptive visual token selection to accelerate VLM inference, with high reproducibility readiness.
2026MASA preprint appears on arXiv: sparse MLLM semantic anchors for wild test-time adaptation, with Wei Huang as last-listed author.
2026AREA preprint appears on arXiv: training-free adaptive evidence selection for knowledge-based VQA in multimodal LLMs.
2026All three catalog papers land in a single year: OpenAlex logs 3 works in 2026 across 12 distinct coauthors, with an average Evolve reproducibility score of 70.

Papers in the Evolve catalog

Frequent co-authors

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

Identity is name-resolved from public OpenAlex bibliographic data; no ORCID is linked, so same-name authors may be conflated.
All metrics come from an OpenAlex snapshot plus Evolve's own reviews of three arXiv preprints; no peer-reviewed venue or citation data is included.
Output and impact dimensions are unscored: h-index and i10-index are 0 and the 2026 preprints have not yet accumulated citations.
Reproducibility reflects Evolve editorial review of three papers only, not a community replication effort.