Siyuan Guo

—

Unclaimed · AI-structured profileOpenAlex

Works at the intersection of tabular foundation models and rigorous AI evaluation, from TabPFN-3.5 to SWE-bench audits

papers here

reviewed by Evolve

first-author

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—

No publication-count data is available in this snapshot, so the output axis is unscored. This dimension would measure total paper volume, venue distribution, and the share of first- or corresponding-author positions over time.

Impact—

Citation data is absent from this profile, so impact is not scored. This axis would measure citation counts, aggregated citation indicators, and the uptake of the reviewed papers by subsequent work.

Focus0/100

The focus axis scores 0, reflecting minimal concentration in a single field: the field distribution splits exactly 50/50 between "AI evaluation and benchmarking (software engineering)" and "machine learning (tabular deep learning and foundation models)". Read alongside the two 2026 papers, this is deliberate breadth rather than drift — both works sit at the interface of building tabular foundation models and stress-testing how AI systems are benchmarked.

Collaboration79/100

Collaboration scores 79, backed by 51 distinct co-authors in the raw record — a wide network for a two-paper reviewed portfolio. The frequent-collaborator list names eleven researchers, including Fengshuo Liu, Ying Liu, Ruize Sun, Lie Luo, Nick Erickson, Léo Grinsztajn, and Kürşat Kaya. Both 2026 papers carry middle and last positions rather than solo authorship, consistent with sustained work inside sizeable, team-based author groups.

Activity—

Publication-cadence data is not yet available, so the activity axis is unscored. This dimension would track yearly output rhythm and the recency of contributions; currently only the two 2026 reviewed papers anchor this profile's timeline.

Reproducibility78/100

Reproducibility averages 78 across the two reviewed papers (reproMean 78 over reproPapers 2). The SWE-bench audit — a statistical re-analysis of 254 submissions using tests such as McNemar's — is rated high reproducibility, while the TabPFN-3.5 report is rated medium, with Evolve noting its headline variants are closed-source. The pattern shows real strength in evaluation rigor, tempered by one release-access caveat.

Research summary

Siyuan Guo is a researcher whose reviewed work spans two tightly connected areas: tabular deep learning and the rigorous evaluation of AI systems. In the Evolve catalog, the portfolio splits evenly between machine learning (tabular deep learning and foundation models) and AI evaluation and benchmarking (software engineering), each accounting for half of the reviewed output. The 2026 TabPFN-3.5 technical report, co-authored in a middle position, claims state-of-the-art results on seven tabular benchmarks through new cell encodings and a doubled-width transformer; Evolve judges the evidence broad and convincing while noting that headline variants remain closed-source. The same year, Guo contributed as last author to an audit of 254 SWE-bench submissions arguing that gaps between top leaderboard entries are statistically inseparable and scaffold-dependent, a study rated highly reproducible. Collaboration is a defining feature: the raw record lists 51 distinct co-authors, a frequent-collaborator list of eleven named researchers, and a collaboration score of 79. Reproducibility averages 78 across the two reviewed papers (one high, one medium readiness). Metrics for output, impact, and publication activity are not yet available in this snapshot, so conclusions about volume or citation standing should await further data.

Collaboration pattern

The raw record lists 51 distinct co-authors, and the frequent-collaborator list names eleven researchers: Fengshuo Liu, Ying Liu, Ruize Sun, Lie Luo, Benjamin Jäger, Nick Erickson, Léo Grinsztajn, Felix Birkel, Klemens Flöge, Oscar Key, and Kürşat Kaya. Both reviewed 2026 papers are multi-author efforts in middle and last positions, consistent with working inside large, team-based collaborations rather than small two-person partnerships.

Research keywords

tabular foundation modelsTabPFNin-context learningtransformer architectureTabArena benchmarknon-i.i.d. evaluationSWE-benchcoding agentsbenchmark saturationleaderboard evaluationMcNemar testmodel-scaffold interaction

Highlights

TabPFN-3.5 claims state of the art on seven tabular benchmarks
The 2026 technical report introduces new cell encodings and a doubled-width transformer, claiming SOTA across seven tabular benchmarks including TabArena-style evaluation. Evolve finds the evidence broad and convincing, though headline variants remain closed-source, and rates reproducibility medium.
SWE-bench's top entries are statistically inseparable
Auditing 254 SWE-bench submissions, the study argues that gaps between top leaderboard entries are statistically inseparable and scaffold-dependent, so the leaderboard can no longer order its top coding agents. Evolve calls it a rigorous, reproducible, mostly convincing evaluation audit.
What to measure instead of the saturated leaderboard
Beyond diagnosing benchmark saturation, the audit proposes alternatives centered on statistical testing (McNemar test) and model-scaffold interaction effects — shifting evaluation from single-number leaderboard rank toward measures that separate model capability from scaffold engineering.

Timeline

2026Co-authors the TabPFN-3.5 technical report in a middle author position, claiming state-of-the-art results on seven tabular benchmarks via new cell encodings and doubled width.
2026Serves as last author on the SWE-bench convergence audit, which reviews 254 submissions and argues top leaderboard gaps are statistically inseparable and scaffold-dependent.
2026The reviewed 2026 output splits evenly between machine learning (tabular deep learning and foundation models) and AI evaluation and benchmarking (software engineering), within a co-author network of 51 distinct collaborators.

Papers in the Evolve catalog

Frequent co-authors

Fengshuo Liu×1 Ying Liu×1 Ruize Sun×1 Lie Luo×1 Siyuan Guo×1 Benjamin J\"ager×1 Nick Erickson×1 L\'eo Grinsztajn×1 Felix Birkel×1 Klemens Fl\"oge×1 Oscar Key×1 K\"ur\c{s}at Kaya×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.

Bibliometric dimensions are computed from public bibliometric data only; institution, country, and ORCID are unverified in this snapshot, and the frequent-coauthor list repeats the researcher's own name, an apparent data artifact.
"Siyuan Guo" is a common name; this profile is name-resolved from public bibliographic records and may conflate homonymous authors.
Output, impact, and activity axes lack underlying data (no publication counts or citation metrics in the snapshot) and are left unscored.
The reproducibility mean of 78 rests on only two reviewed papers, so it should be read as indicative rather than stable.