Zhejiang University researcher on LLM post-training: data-efficient Text-to-SQL and interaction-based spatial reasoning
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) 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: publication-volume data is unavailable in this snapshot. This axis would measure total papers per year, authorship positions, and venue spread across Shen's full bibliographic record, beyond the two cataloged 2026 arXiv papers.
Impact—
Not yet scored: citation data is unavailable in this snapshot. This axis would measure citation counts, aggregate influence, and uptake of Shen's work; both cataloged papers are 2026 preprints, too recent for stable citation signals.
Focus0/100
The focus score is 0/100, reflecting a field distribution split exactly evenly in the raw data: 50% machine learning (multimodal vision-language models) and 50% natural language processing (Text-to-SQL / semantic parsing). Under the radar's concentration measure, an even two-way split yields no dominant specialty. The two cataloged papers embody this: LIMIT targets Text-to-SQL instruction tuning while Spatial-Interactor trains VLM spatial reasoning from interaction trajectories.
Collaboration56/100
The collaboration score is 56/100, computed from 16 distinct coauthors in the snapshot. Recurring names include Kaixiang Yao, Xu Wang, Wenqi Zhang, and Daniel Dahlmeier. On both cataloged 2026 papers Shen holds a middle authorship position without corresponding-author status, a pattern consistent with contributing inside larger teams — one NLP (Text-to-SQL) effort and one multimodal vision-language effort — mirroring the even 50/50 field split.
Activity—
Not yet scored: longitudinal activity data is insufficient. This axis would measure publication cadence and year-over-year research continuity, but the snapshot covers only two papers, both from 2026, so no ongoing-activity trend can be established.
Reproducibility55/100
The reproducibility score is 55/100, the mean over 2 reviewed papers (reproPapers = 2), and both carry a 'medium' reproducibility readiness rating in Evolve's reviews. For LIMIT, the pipeline is well ablated but schema-overlap and statistical-significance concerns remain; for Spatial-Interactor, the reported interaction-trajectory training is broadly effective, yet the paper's own ablation attributes most gains to generic spatial-QA SFT, a caveat for anyone reimplementing it.
Research summary
Yongliang Shen is a researcher at Zhejiang University whose cataloged work spans two complementary branches of modern LLM post-training. In natural language processing, Shen contributed to LIMIT (2026), a Text-to-SQL instruction-tuning study showing that 796 carefully selected samples reach 69.1% BIRD EX with Qwen3-8B; Evolve's review credits a well-ablated pipeline while flagging schema-overlap and statistical-significance concerns. In multimodal learning, Shen contributed to Spatial-Interactor (2026), which trains vision-language models for spatial reasoning from interaction trajectories — the LSI-108K dataset with SFT plus on-policy distillation — though its own ablation finds generic spatial-QA SFT drives most gains. The bibliometric snapshot splits Shen's field distribution evenly (50/50) between multimodal vision-language work and Text-to-SQL/semantic parsing, yielding a focus concentration score of 0/100. The snapshot records 16 distinct coauthors (collaboration score 56/100), with Wenqi Zhang, Xu Wang, and Kaixiang Yao among recurring collaborators, and a reproducibility mean of 55/100 across the two reviewed papers, both rated medium reproducibility readiness. On both cataloged papers Shen appears in a middle authorship position rather than as corresponding author.
Collaboration pattern
The snapshot records 16 distinct coauthors (collaboration score 56/100). The frequent-coauthor list — including Kaixiang Yao, Xu Wang, Miao Pan, Hu Xiyue, Weishi Wang, Jintao Chen, Xuhong Zhang, Wenqi Zhang, Haoyuan Ma, Hengwei Liu, Linjuan Wu, and Daniel Dahlmeier — spans both research tracks. On both 2026 cataloged papers Shen appears in a middle authorship position without corresponding-author status, consistent with contributing inside larger teams rather than leading them.
Research keywords
Highlights
LIMIT: 796 curated samples hit 69.1% BIRD EX on Qwen3-8B
Spatial-Interactor trains VLM spatial reasoning from LSI-108K interaction trajectories
Middle-author contributions across two 2026 preprints in distinct subfields
Both cataloged papers list Shen in a middle authorship position without corresponding-author status — one in Text-to-SQL instruction tuning (with genetic-algorithm-driven data selection and LLM-as-judge filtering), one in multimodal VLM training (with GRPO and on-policy distillation) — matching the snapshot's even 50/50 field distribution across 16 coauthors.
Timeline
- 2026Co-authored Spatial-Interactor (arXiv), an interaction-trajectory approach (LSI-108K; SFT plus on-policy distillation) to spatial reasoning for vision-language models, in a middle authorship position.
- 2026Co-authored LIMIT (arXiv), showing that 796 curated instruction-tuning samples reach 69.1% BIRD EX on Qwen3-8B for Text-to-SQL, in a middle authorship position.
- 2026Both papers entered the Evolve review catalog with medium reproducibility readiness; the snapshot records 16 distinct coauthors and an even 50/50 split across the two research fields.
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 scores are computed from a limited snapshot (OpenAlex-style data); output, impact, and activity axes lack the underlying data and are unscored.
- Only two papers, both 2026 arXiv preprints reviewed by Evolve, inform this profile; no citation counts, h-index, career history, or award information is included.
- No ORCID or country is present in the source data; the profile is name-resolved from public bibliographic data via the listed Zhejiang University affiliation.
- The reproducibility mean (55/100) rests on only 2 reviewed papers; treat the score as indicative rather than a stable estimate.