Xu Wang

Gunma University

Unclaimed · AI-structured profileOpenAlex

VLM spatial reasoning meets SBI-based social-simulator validation; early-stage profile from two 2026 arXiv preprints

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-13) 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-output data is insufficient for this profile. This axis would measure total paper count, venue tier and authorship role over time; it can be computed once a fuller publication record is linked to this identity.

Impact—

Not yet scored — citation and uptake data (citation counts, h-index, citing works) is unavailable in the current snapshot. This axis would measure how widely the two reviewed 2026 papers and any earlier work have been cited and adopted.

Focus0/100

Scores 0/100 — the field distribution is an even 50/50 split between machine learning (multimodal vision-language models) and computational social science (agent-based modeling and simulation-based inference), so no dominant field emerges. With only two reviewed papers, one per field, this may be genuine breadth or an unsettled agenda; distinguishing the two needs a longer publication record.

Collaboration54/100

Scores 54/100. The record lists 14 distinct coauthors, with 12 recurring collaborators — including Kaixiang Yao, Yongliang Shen and Masayuki Goto — spanning the machine-learning and social-simulation communities. On both reviewed papers Xu Wang sits in a middle authorship position without corresponding-author credit, consistent with a contributing member of mid-sized teams rather than an independent lead; the moderate score reflects breadth across two communities, not yet a dense, self-led network.

Activity—

Not yet scored — the data needed to measure publishing cadence (per-year output, submission dates, career span) is insufficient. Both reviewed papers carry 2026 dates, so this axis would track output volume and consistency across recent years once more records exist.

Reproducibility55/100

Scores 55/100, the mean reproducibility rating across 2 papers carrying reproducibility data; both are individually rated medium readiness by Evolve. The spatial-reasoning paper leans on a sizeable training corpus (LSI-108K) and a multi-stage SFT-plus-on-policy-distillation pipeline, while the social-simulation paper depends on SBI calibration and audit code; both pipelines are non-trivial to re-run, consistent with medium rather than high ratings.

Research summary

Xu Wang, affiliated with Gunma University, is an early-stage researcher profiled here from public bibliographic data; the name is common, so this profile is name-resolved rather than biography-verified. The Evolve catalog contains two reviewed papers, both 2026 arXiv preprints with Xu Wang in a middle authorship position, and they split evenly across two fields: multimodal vision-language models and computational social science. In machine learning, Spatial-Interactor introduces interaction-trajectory training (the LSI-108K corpus, with SFT plus on-policy distillation under GRPO) to lift spatial reasoning in vision-language models; Evolve's verdict credits broad gains while noting the paper's own ablation attributes most improvement to generic spatial-QA SFT. In computational social science, the adequacy-testing paper adds simulation-based-inference calibration — amortized neural posterior estimation, posterior-predictive checking, repair and a held-out audit — to generative social simulators, honestly reporting that the tested simulator remains inadequate against emergent network structure. The record lists 14 distinct coauthors and 12 frequent collaborators spanning both communities, with a collaboration score of 54 and reproducibility score of 55 (mean over 2 papers, each rated medium readiness). Output, impact and activity are unscored for lack of data. This is a thin, two-paper evidence base.

Collaboration pattern

The record lists 14 distinct coauthors, with a recurring core of 12 frequent collaborators including Kaixiang Yao, Miao Pan, Hu Xiyue, Weishi Wang, Daniel Dahlmeier, Yongliang Shen, Wenqi Zhang and Masayuki Goto. On both reviewed papers Xu Wang is a middle author and not the corresponding author, consistent with a contributor role inside mid-sized teams that span machine learning and computational social science; the collaboration score is 54/100.

Research keywords

agent-based modellingsimulation-based inferenceamortized neural posterior estimationposterior-predictive checkingmodel adequacy and misspecificationgenerative social simulationspatial reasoningvision-language modelsinteraction trajectoriesstate transitionson-policy distillationgroup relative policy optimization (GRPO)

Highlights

SBI-based adequacy testing for generative social simulators
This 2026 arXiv paper proposes calibrating generative social simulators against emergent network structure using simulation-based inference — amortized neural posterior estimation, posterior-predictive checking, repair and a held-out audit — rather than presuming simulator adequacy.
Honest negative result: the tested simulator stays inadequate
Evolve's verdict credits the paper as rigorous and honest: even with the repair loop and held-out audit, the tested generative social simulator still does not match emergent network structure, making the work a template for adequacy testing rather than a claimed success.
Interaction-trajectory training lifts VLM spatial reasoning
Spatial-Interactor trains vision-language models on interaction trajectories with the observable physical world (LSI-108K; SFT plus on-policy distillation under GRPO), and Evolve's verdict notes it lifts VLM spatial reasoning broadly.
Self-flagged caveat in Spatial-Interactor's ablation
Evolve's review highlights that the paper's own ablation shows generic spatial-QA supervised fine-tuning drives most of the gains, tempering the claimed benefit of the interaction-trajectory and on-policy-distillation components.

Timeline

2026Published (arXiv, middle authorship) "Testing, not presuming, adequacy", introducing SBI-based adequacy testing, repair and a held-out audit for generative social simulators.
2026Published (arXiv, middle authorship) "Spatial-Interactor", proposing interaction-trajectory training for vision-language model spatial reasoning (LSI-108K; SFT plus on-policy distillation).
2026Both papers entered the Evolve review catalog, producing a 50/50 two-field split between vision-language models and computational social science, 14 distinct coauthors, and a mean reproducibility score of 55.

Papers in the Evolve catalog

Frequent co-authors

Kaixiang Yao×1 Miao Pan×1 Hu Xiyue×1 Weishi Wang×1 Daniel Dahlmeier×1 Jintao Chen×1 Yongliang Shen×1 Xuhong Zhang×1 Wenqi Zhang×1 Tengfei Shao×1 Chao Li×1 Masayuki Goto×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 from public bibliographic data: "Xu Wang" is a common name, and the link to Gunma University should be treated as provisional without an ORCID.
All evidence comes from two Evolve-reviewed 2026 arXiv preprints plus a bibliometric snapshot; no citation counts, h-index or venue-acceptance data were available.
Output, impact and activity dimensions are unscored because the underlying bibliometric data is insufficient; no estimates were substituted.
Collaboration and reproducibility scores are computed only from the reviewed-paper sample (2 papers, 14 distinct coauthors), not from a full publication history.