Multimodal agent researcher: scientific-figure reconstruction benchmarks and efficient GUI grounding with VLMs
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-20) 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—
Output metrics are not available in the current snapshot: this axis would measure total publication count, venue distribution, and first- versus middle-author share over time. Until publication-volume data is ingested, no score can be shown for this researcher.
Impact—
Impact cannot be scored yet: citation counts, citation velocity, and citing venues are absent from the available data. This axis would track how widely ReFigBench and RankGround are adopted by other benchmark, evaluation, and GUI-agent papers.
Focus0/100
The focus score of 0 reflects an even 50/50 field distribution: machine learning (multimodal agents and evaluation) and computer vision (GUI agents and vision-language models) each account for 0.5. This reads as breadth across two halves of one agenda — the reviewed work builds and stress-tests multimodal agents, whether reconstructing scientific figures or grounding GUI interfaces.
Collaboration56/100
A collaboration score of 56 rests on 16 distinct coauthors across the reviewed work. The recurring co-occurrence of Liyang Fan and Yitai Li suggests a stable core team, extended by rotating contributors such as Chi Wei, Guhong Chen, and Chenghao Sun. Middle-author positioning on both catalog papers fits a contributor role inside larger groups rather than solitary work.
Activity—
Activity data is insufficient: per-year publication rate and year-over-year momentum are not present in the snapshot. This axis would measure publishing cadence and consistency across years, which cannot be estimated from the two reviewed 2026 papers alone.
Reproducibility55/100
Reproducibility averages 55 across the 2 papers with recorded readiness, both rated medium. Evolve's verdicts note methodological care — the paired harness-versus-workflow study in ReFigBench and honest analysis in RankGround — while flagging judge-family caveats and a +5.5% headline needing scrutiny. The score signals solid but not yet turnkey replication.
Research summary
Xinping Bi is a researcher working at the junction of multimodal agent evaluation and vision-language systems for graphical interfaces. Evolve's catalog includes two 2026 arXiv papers with Xinping Bi in a middle author position on each. The first, ReFigBench, introduces a 1,000-figure benchmark for reconstructing scientific figures as editable PowerPoint artifacts; a paired study shows that harness and workflow jointly decide agent quality, with judge-family caveats noted. The second, RankGround, tackles high-resolution GUI grounding through lightweight reranker-guided crop selection, needing only a single vision-language model call; Evolve judged the results strong and honestly analyzed while flagging that the +5.5% headline needs scrutiny. Together the two papers sketch a coherent agenda: making multimodal agents measurable and efficient, whether they edit figures or operate interfaces. Bibliometrically, the reviewed work splits evenly between machine learning (multimodal agents and evaluation) and computer vision (GUI agents and vision-language models), with 16 distinct coauthors and recurring collaborators. Reproducibility readiness is rated medium on both papers. Both appeared in 2026, marking an early, active phase of this profile.
Collaboration pattern
Sixteen distinct coauthors appear across the reviewed work, matching the moderate collaboration score of 56. The co-occurrence list repeats Liyang Fan and Yitai Li, pointing to a stable core duo, alongside contributors such as Chi Wei, Guhong Chen, Chenghao Sun, Haoxiang Yang, Qingwen Li, Kai Yan, Hong Li, and Bo Li. Middle-author positioning on both papers suggests a contributor within a larger, team-based research group rather than a solo line of work.
Research keywords
Highlights
ReFigBench: a 1,000-figure test of rebuilding science figures as editable slides
RankGround: cheaper high-resolution GUI grounding via reranker-guided crops
A consistent evaluation-first signature
Both catalog papers stress-test how multimodal systems are measured: one benchmarking agent workflows for figure reconstruction, the other auditing its own headline gain — with medium reproducibility readiness on each.
Timeline
- 2026ReFigBench posted to arXiv, introducing a 1,000-figure benchmark for reconstructing scientific figures as editable PowerPoint artifacts.
- 2026RankGround posted to arXiv, presenting lightweight reranker-guided crop selection for efficient high-resolution GUI grounding.
- 2026Both Evolve-reviewed papers land in the same year, establishing an evaluation-first agenda spanning multimodal agent benchmarking and GUI grounding.
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.
- Bibliometric dimensions are computed from an OpenAlex-derived snapshot only; citation counts, h-index, and publication-volume data are absent, leaving output, impact, and activity unscored.
- The profile is name-resolved from public bibliographic data; institution, country, and ORCID are unavailable, so identity matching for this name carries some uncertainty.
- Qualitative assessments rest on only two Evolve-reviewed papers, both 2026 arXiv preprints not yet peer-reviewed.
- The reproducibility score of 55 is averaged over just 2 papers, both rated medium readiness, so it may not generalize.