LLM agent researcher at Tsinghua building verifiable computer-use environments and mapping recursive self-improvement
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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: the snapshot provides no publication-volume, venue, or productivity metrics beyond the two reviewed papers. This axis would measure total research output and its distribution across venues and years once fuller bibliographic records are available.
Impact—
Not yet scored: no citation counts, h-index, or downstream-adoption signals are present in the input data. This axis would measure citation-based influence and how far the benchmark and the autonomy taxonomy are taken up by other agent-research groups.
Focus0/100
The focus axis is recorded as 0/100 in this snapshot, but the underlying raw field distribution tells a different story: cataloged output splits 0.5 to 'machine learning (LLM agents and self-improving AI systems)' and 0.5 to 'machine learning / LLM agents' — two labels for effectively one topic. The zero likely reflects label fragmentation rather than genuine thematic spread.
Collaboration84/100
The collaboration axis scores 84/100, grounded in 68 distinct coauthors in the bibliometric snapshot and a recurring core of 12 frequent collaborators, including Yi Duan, Ying Liu, Zirui Tang, Haodong Chen, Jun Zhou, and Yumou Liu. This co-author breadth indicates large, lab-scale team projects — the pattern expected for benchmark construction — rather than small dual-author work.
Activity—
Not yet scored: the snapshot lacks per-year publication frequency. This axis would measure research cadence and continuity over time; only 2026-dated papers are currently visible, so no multi-year activity trend can be established.
Reproducibility60/100
Reproducibility scores 60/100 (reproMean 60 over reproPapers 2). The two reviewed papers sit at opposite ends: RecreationWorld is rated high readiness — five platforms, verifiable rewards, and an SFT pipeline make it directly reusable — while the recursive self-improvement survey is rated low because its taxonomy and metrics are conceptual. Empirical artifact-building thus reproduces far better than survey work.
Research summary
Bowen Zhou is a researcher at Tsinghua University whose cataloged work centers on large language model agents — both the environments they act in and the question of how far they can improve themselves. Two 2026 arXiv papers frame this agenda. The first, RecreationWorld, introduces a five-platform recreation benchmark with verifiable rewards and an SFT pipeline for hybrid GUI/coding computer-use agents; Evolve rates its reproducibility readiness high while noting that transfer gains lack statistical controls. The second, 'The Last AI Built by Humans,' surveys recursive self-improvement and proposes a five-level autonomy taxonomy supported by HCI analysis and industry cases; Evolve finds the taxonomy useful but the metrics ad hoc, rating reproducibility low. Together they span agent evaluation infrastructure and long-horizon autonomy — one builds verifiable environments, the other maps the trajectory toward self-improving systems. The bibliometric snapshot records 68 distinct coauthors and 12 frequent collaborators, yielding a collaboration score of 84, and a field distribution split evenly across two LLM-agent labels, indicating a tightly focused topic profile. Output, impact, and activity axes remain unscored because citation and volume data are not yet available; this should be read as an early-stage, catalog-based profile rather than a complete bibliometric record.
Collaboration pattern
Evolve records 68 distinct coauthors and a collaboration score of 84, with a recurring core of 12 frequent collaborators including Yi Duan, Ying Liu, Zirui Tang, Haodong Chen, Jun Zhou, and Yumou Liu. This breadth points to large, lab-scale team projects consistent with benchmark construction, though all cataloged coauthorship currently derives from the two 2026 arXiv papers.
Research keywords
Highlights
Five-platform verifiable benchmark for hybrid computer-use agents
Five-level autonomy taxonomy for recursive self-improvement
Timeline
- 2026Published RecreationWorld on arXiv in a last-author position: a five-platform recreation benchmark with verifiable rewards plus an SFT pipeline for hybrid GUI/coding computer-use agents.
- 2026Co-authored 'The Last AI Built by Humans,' a survey proposing a five-level autonomy taxonomy for recursive self-improvement with HCI analysis and industry cases.
- 2026Both papers entered the Evolve review catalog, with reproducibility readiness rated high for RecreationWorld and low for the self-improvement survey.
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 name-resolved from public bibliographic data: 'Bowen Zhou' is a common name and no ORCID is linked, so identity convergence with the Tsinghua University affiliation is uncertain.
- Bibliometric dimensions come from a sparse snapshot: output, impact, and activity lack sufficient data and are left unscored, not zero.
- The reproducibility score (60) rests on only 2 reviewed papers and the collaboration figures on their author lists, so the sample size is small.
- The focus axis is recorded as 0 while the field distribution shows output concentrated in one topic area, which likely reflects a label-normalization artifact rather than dispersed research.