Xu Wang

—

Unclaimed · AI-structured profileOpenAlex

Benchmark and validation methodology across LLM unlearning and generative social simulation

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: publication counts, venue types, and authorship volume are absent from this snapshot beyond the two 2026 arXiv papers. This axis would measure total output volume and the share of first or corresponding-author work over time.

Impact—

Not yet scored: citation counts, h-index-type indicators, and referencing works are missing from the input data. This axis would track uptake and influence of the cataloged benchmarks and methods across the wider literature.

Focus0/100

Focus scores 0 because the cataloged output splits exactly 50/50 between two distinct areas: machine learning on LLM privacy and unlearning, and computational social science built on agent-based modeling with simulation-based inference. One 2026 paper sits in each field. What unites them is method — benchmark construction and adequacy testing — rather than topic, so topical concentration is genuinely low on this metric rather than an artifact.

Collaboration46/100

Collaboration scores 46, derived from 9 distinct co-authors in the bibliometric record. Xu Wang holds non-corresponding positions on both 2026 papers — middle author on the adequacy-testing study and last author on K-Bench. Recurring names include Guangsheng Yu, Yanna Jiang, Qin Wang, Baihe Ma, Tengfei Shao, Chao Li, and Masayuki Goto, consistent with small-to-medium team science rather than solo authorship.

Activity—

Not yet scored: year-by-year publication cadence cannot be computed from a snapshot containing only two 2026 papers. This axis would measure recent output pace, preprint frequency, and continuity of research activity across consecutive years.

Reproducibility78/100

Reproducibility is the strongest measured axis at 78, the mean across 2 scored papers. K-Bench rates high readiness — rigorously built and explicitly pre-registered — while the SBI-based adequacy paper rates medium. Both works foreground verifiable artifacts: held-out audits, posterior-predictive checks, and pre-registered protocols, pointing to an evaluation-first working style that supports independent replication.

Research summary

Xu Wang is an early-stage researcher whose cataloged work splits evenly — 50/50 in the recorded field distribution — between machine learning on LLM privacy and unlearning, and computational social science built on agent-based modeling and simulation-based inference. Both reviewed papers appeared on arXiv in 2026 and share one stance: build the benchmark or the test first, then report what fails. K-Bench, where Xu Wang is last author, is a pre-registered benchmark showing that agentic deployments — ReAct-style agents and chain-of-thought leakage among them — defeat single-channel unlearning certificates; Evolve's review finds the headline negative holds even though method rankings shift across base models. The adequacy paper, where Xu Wang is a middle author, adds SBI-based adequacy testing, repair, and a held-out audit to generative social simulators and reports, honestly, that the tested simulator stays inadequate. Reproducibility is the strongest measured axis at 78, the mean across two papers rated high and medium readiness. Collaboration is moderate at 46 across 9 distinct co-authors, with non-corresponding roles on both papers. Output, impact, and activity lack underlying metrics and are left unscored. No affiliation, country, or ORCID is available; this profile is name-resolved from public bibliographic data.

Collaboration pattern

The bibliometric record shows 9 distinct co-authors and a moderate collaboration score of 46. Xu Wang is neither first nor corresponding author on either 2026 paper — middle author on the adequacy study, last author on K-Bench. Recurring names include Guangsheng Yu, Yanna Jiang, Qin Wang, Baihe Ma, Tengfei Shao, Chao Li, and Masayuki Goto. Which partner maps to which paper is not specified in the snapshot.

Research keywords

LLM unlearningMachine unlearning benchmarksAgentic evaluationReAct agentsPrivacy leakageChain-of-thought leakageAgent-based modellingSimulation-based inferenceAmortized neural posterior estimationPosterior-predictive checkingModel adequacy and misspecificationGenerative social simulation

Highlights

K-Bench: agentic deployments defeat single-channel unlearning certificates
A pre-registered benchmark for LLM unlearning in agentic deployments, built rigorously and rated high reproducibility readiness. Evolve's review finds the headline negative holds — no published method removes the secret across all channels — though method rankings shift across base models.
Testing, not presuming, adequacy in generative social simulators
Adds simulation-based-inference adequacy testing, repair, and a held-out audit to social-simulator validation against emergent network structure. The review calls it rigorous and honest — while the paper itself reports that the tested simulator stays inadequate. Rated medium reproducibility readiness.
Negative results kept on the record
Both cataloged papers publish their headline failures rather than burying them: no published method removes the secret across all channels in K-Bench, and the social simulator remains inadequate despite repair attempts. This honesty pairs with a reproducibility mean of 78 across the two scored papers.

Timeline

2026Co-authors K-Bench, a pre-registered benchmark for LLM unlearning in agentic deployments (arXiv; last author, non-corresponding).
2026Contributes to the adequacy-testing study calibrating generative social simulators against emergent network structure via simulation-based inference (arXiv; middle author, non-corresponding).
2026Cataloged output splits evenly across the two fields, with a reproducibility mean of 78 over the two papers and 9 distinct co-authors on record.

Papers in the Evolve catalog

Frequent co-authors

Guangsheng Yu×1 Yanna Jiang×1 Qin Wang×1 Baihe Ma×1 Tengfei Shao×1 Chao Li×1 Xu Wang×1 Masayuki Goto×1 Tengfei Shao×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.

Radar scores and co-author counts come from public bibliometric data only; no affiliation, country, or ORCID was available, and 'Xu Wang' is a common name, so identity is resolved from bibliographic records rather than a verified author ID (Chinese-character name unconfirmed).
Output, impact, and activity axes are unscored because citation, volume, and cadence metrics were absent from the input snapshot.
Qualitative verdicts rest on only two reviewed papers, both 2026 arXiv preprints; no journal publications, grants, or awards are in scope.
The co-author-to-paper mapping is not provided; recurring names are aggregated at profile level, not per paper.