Harbin Institute of Technology researcher working on LLM agent security auditing and memory for closed-loop VLA agents
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-17) 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—
Insufficient data: this snapshot provides no publication counts, venue tiers, or per-year output totals, so output is unscored. The axis would measure total indexed publications and annual production volume once bibliographic coverage for this author is complete.
Impact—
Insufficient data: no citation counts, h-index, or downstream-adoption signals exist in the current snapshot, so impact is unscored. The axis would measure citation volume and uptake of the agent-security tooling across the literature.
Focus100/100
The focus axis scores 100/100: the raw field distribution places the one classified catalog paper entirely in AI security (LLM agent safety), with no competing fields. Both 2026 arXiv entries — security auditing for mobile GUI agents and achievement-grounded memory for closed-loop agents with frozen VLA policies — sit on a single tightly defined agenda of making LLM-driven agents dependable, indicating exceptional early concentration rather than scattered interests.
Collaboration46/100
The collaboration axis scores 46/100, based on 9 distinct coauthors in the raw data. The named frequent collaborators — Changyue Jiang, Jiayi Wang, Jiarun Dai, Geng Hong, Xudong Pan, Hongbo Gao, Zeyu Ni, Siyu Xu, and Ruifeng Li — form a single dense cluster. On both cataloged papers Xin Wen appears in a middle authorship position without corresponding status, consistent with a contributing member of a moderately sized, stable team rather than an independent group lead so far.
Activity—
Insufficient data: the snapshot contains only two 2026 arXiv entries with no full publication history, so activity is unscored. The axis would measure recent publication cadence and year-over-year research continuity once a longer indexed record is available.
Reproducibility55/100
Reproducibility scores 55/100, computed as reproMean 55 over just one assessed paper (reproPapers 1), so the axis rests on a thin base. The assessed work, MATE, is rated medium repro-readiness: it trains an audit model on 140K synthetic trajectories, but Evolve flags that labels produced by LLMs risk circularity, which complicates independent verification and replication of the claimed audit performance.
Research summary
Xin Wen is a researcher at Harbin Institute of Technology whose visible catalog record centers on the dependability of LLM-driven agents. Two 2026 arXiv papers define the current snapshot. The first, MATE, presents a policy-aware security auditing framework for mobile agents: a lightweight auditor trained via synthesis-driven trajectory learning on 140K synthetic trajectories, which checks agent traces against policies; Evolve's review credits strong results while cautioning that labels produced by LLMs risk circularity, and rates repro-readiness as medium. The second, AGM, proposes achievement-grounded memory for closed-loop agents built on frozen vision-language-action policies, extending the same agenda from auditing agent behavior to structuring agent memory. The radar reflects an early-stage but sharply concentrated profile: focus scores 100/100 with the classified field distribution mapping entirely to AI security (LLM agent safety), reproducibility scores 55/100 from a single assessed paper, and collaboration scores 46/100 across 9 distinct coauthors. Xin Wen appears in middle authorship positions without corresponding status on both papers. Output, impact, and activity axes lack underlying bibliometric coverage and are left unscored.
Collaboration pattern
The raw record shows 9 distinct coauthors, and the frequent-collaborator list — Changyue Jiang, Jiayi Wang, Jiarun Dai, Geng Hong, Xudong Pan, Hongbo Gao, Zeyu Ni, Siyu Xu, Ruifeng Li — forms one stable cluster. On both 2026 papers Xin Wen sits in middle authorship positions without corresponding status, consistent with a contributing member of a moderately sized team; the collaboration score of 46/100 reflects moderate rather than expansive network breadth.
Research keywords
Highlights
MATE: policy-aware security auditing for mobile agents
Scaling audit training data through synthesis
Achievement-grounded memory for closed-loop VLA agents
A second 2026 arXiv paper, AGM, proposes achievement-grounded memory for closed-loop agents running frozen vision-language-action policies, extending the agent-reliability agenda from auditing behavior to structuring memory; Evolve has not yet reviewed it, so no verdict or repro-readiness is available.
Timeline
- 2026MATE: Policy-Aware Security Auditing for Mobile Agents appears on arXiv, with Xin Wen in a middle, non-corresponding authorship position.
- 2026Evolve reviews MATE, rating repro-readiness medium and cautioning that LLM-generated labels risk circularity.
- 2026AGM: Achievement-Grounded Memory for Closed-Loop Agents with Frozen VLA Policies appears on arXiv, again with middle authorship and no corresponding role.
Papers in the Evolve catalog
Frequent co-authors
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.
- Profile is name-resolved from public bibliographic data: no ORCID is provided and 'Xin Wen' is a common name, so identity confidence is limited.
- Bibliometrics are drawn from OpenAlex only; output, impact, and activity axes lack underlying data and are unscored.
- Only the MATE paper carries a full Evolve review; the AGM entry has no verdict, keywords, or repro-readiness assessment yet.
- Country is not recorded in the source data; the institutional affiliation (Harbin Institute of Technology) is the only location signal.