LLM-agent researcher at Beihang University: agent memory, cross-device GUI benchmarks, and embodied evaluation
3
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
1
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 to score: the OpenAlex snapshot shows only thin cumulative output records (h-index 1, i10-index 0). This axis would measure total works per year, venue spread, and first/corresponding-author share.
Impact—
Not yet scoreable: citation records in the snapshot are too sparse (i10-index 0), so citation-based impact cannot be estimated. This axis would measure total citations, citation trajectory, and uptake of the benchmark papers.
Focus0/100
The focus score is 0, reflecting an even 50/50 split in the field distribution between NLP (LLM-based agents) and AI (GUI agents / human-computer interaction). Five OpenAlex topics — information retrieval and search behavior (3), AI-based problem solving and planning (3), topic modeling (2), recommender systems (2), reinforcement learning in robotics (2) — confirm breadth across agent-related subfields, though all three catalog papers share an agents-and-evaluation core.
Collaboration62/100
Collaboration scores 62, backed by 21 distinct coauthors in the recent three-year window. Frequent collaborators include Wanxiang Che, Yuhang Guo, Yuli Qiu, Yutong Li, Wei Su, Heyan Huang, and Haifeng Wang, spanning NLP and interaction-oriented groups. On all three reviewed papers Liu is a middle author and not corresponding, suggesting a contributor role inside larger team-driven projects rather than solo-led studies.
Activity67/100
Activity scores 67 on a steep recent ramp: 9 works in the last three years, with 8 indexed in 2026 versus 1 in 2025, and all three Evolve-catalog papers dated 2026. This indicates a burst of arXiv-first output within a single year; the observation window is short, so a sustained pace beyond 2026 cannot yet be confirmed.
Reproducibility55/100
Reproducibility averages 55 across 2 reviewed papers. Evolve's verdicts converge on a pattern: LifeMem's 10-environment study reports no significance tests or variance; JarvisGUI's near-zero multi-device scores may partly reflect the shared planner harness; the household-device benchmark relies on single-run evaluations, synthetic renders, and a 3x Gemini outlier. All three catalog papers are rated medium reproReadiness, so artifacts exist but statistical rigor is the recurring gap.
Research summary
Zeming Liu is an early-career researcher affiliated with Beihang University whose recent work concentrates on large language model (LLM) agents: how they remember and reuse experience, operate across devices, and explore physical environments. The Evolve catalog holds three 2026 arXiv preprints with Liu as a middle author: LifeMem on lifelong experience reuse, JarvisGUI on cross-device GUI agents with dynamic task composition, and a benchmark for generalizable visually grounded exploration of household devices. OpenAlex topics mirror this spread — information retrieval and search behavior (3), AI-based problem solving and planning (3), topic modeling (2), recommender systems (2), and reinforcement learning in robotics (2) — and the field distribution splits evenly between NLP (LLM-based agents) and AI (GUI agents / human-computer interaction). Activity is concentrated and recent: 9 works in the last three years, 8 of them indexed in 2026, alongside 21 distinct coauthors. Evolve's reproducibility reviews, covering two papers with a mean readiness of 55, repeatedly flag missing significance tests, single-run evaluations, and possible harness artifacts — a consistent theme in otherwise ambitious benchmark work. Output and impact axes remain unscored because cumulative citation records are still too thin (h-index 1, i10-index 0).
Collaboration pattern
With 21 distinct coauthors in the last three years and middle-author, non-corresponding positions on all three reviewed papers, Liu works in team-based settings. Recurring collaborators include Wanxiang Che, Yuhang Guo, Heyan Huang, and Haifeng Wang, alongside Yuli Qiu, Yutong Li, Wei Su, Yuang Guo, Zixiang Chen, Yuheng Lu, Zihao Cheng, and Jizeng Bai — a network spanning NLP, GUI-agent, and embodied-evaluation work.
Research keywords
Highlights
LifeMem: distilling reusable agent skills from clustered trajectories
JarvisGUI: first cross-device GUI agent benchmark across three OSes
Manual-free benchmark for visually grounded household device exploration
Timeline
- 2026LifeMem preprint appears: lifelong experience reuse for LLM agents via workflow clustering of agent trajectories.
- 2026JarvisGUI preprint appears: a cross-device GUI agent benchmark spanning three operating systems with typed task auto-composition.
- 2026Household-device exploration benchmark preprint appears, extending the evaluation agenda to embodied visual grounding.
- 2026Year of concentration: 8 of the 9 recent three-year works are indexed in 2026, built with 21 distinct coauthors.
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.
- Bibliometrics come from an OpenAlex snapshot only (identity: openalex); no ORCID was provided, so this profile is name-resolved and 'Zeming Liu' is a common name.
- Output and impact axes are unscored because cumulative citation records in the snapshot are too thin (h-index 1, i10-index 0).
- Provided metadata lists institution Beihang University with country 'US'; this mismatch was left unresolved rather than corrected by guesswork.
- Reproducibility draws on Evolve reviews of only 2 papers, and all three catalog entries are arXiv preprints without peer-review status.