NLP researcher in data-efficient LLM post-training: Text-to-SQL instruction tuning and rubric-rewarded plan generation
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
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—
Output is unscored: the snapshot contains no publication-count or productivity measure for this name. This axis would aggregate cataloged papers, venues, and yearly volume; it needs a broader bibliographic crawl before any rating can be made.
Impact—
Impact is unscored: no citation counts, h-index, or downstream-adoption signals appear in the input. This axis would measure citations and scholarly influence of the work; without those figures no estimate is offered.
Focus100/100
The focus axis scores 100/100, the maximum in this snapshot. The raw field distribution records exactly one field — natural language processing (text-to-SQL / semantic parsing) — and both 2026 catalog papers sit inside NLP: Text-to-SQL instruction tuning on one side, LLM post-training with rubric rewards for research-plan generation on the other. Within the coverage window this is a tightly concentrated, data-centric post-training agenda rather than a spread across disciplines.
Collaboration54/100
Collaboration scores 54/100, a mid-range value. The snapshot records 14 distinct coauthors, and the frequent-coauthor list — after removing the researcher's own duplicated entry — clusters around recurring names such as Weiming Lu, Yuhan Wang, Zhengxi Lu, Yuchen Yan, Wenqi Zhang, Jun Xiao, and Yueting Zhuang. Both 2026 papers are multi-authored team outputs, with a middle-author slot on LIMIT and the last-author slot on PaperGym, neither as corresponding author.
Activity—
Activity is unscored: only two papers are present, both dated 2026, so no multi-year publishing cadence can be computed. This axis would track publication frequency and recent output over time; a longer coverage window is required.
Reproducibility55/100
Reproducibility scores 55/100, averaged over exactly one paper (reproPapers = 1), so the axis rests on a thin base. Both catalog papers carry a medium Evolve repro-readiness rating. Evolve's verdicts temper confidence: LIMIT is described as well ablated but with unresolved schema-overlap and significance concerns, while PaperGym's gains rest on LLM-judge evaluations without significance testing. Replication should prioritize releasing the selection pipelines and rubric data.
Research summary
Yongliang Shen is a natural language processing researcher whose Evolve-visible work centers on data-efficient post-training of large language models. The catalog holds two 2026 arXiv preprints. The first, LIMIT: Less Is More for Instruction Tuning in Text-to-SQL (middle author), reports a strong data-efficiency result: 796 curated samples yield 69.1% execution accuracy on BIRD with Qwen3-8B, with selection driven by a genetic algorithm and LLM-as-judge scoring; Evolve's review credits a well-ablated pipeline while flagging schema-overlap and significance concerns. The second, PaperGym (last author), targets research-plan generation with a rubric-as-reward scheme: rubric data is decoupled with a reported 3.7% leakage rate, and OPSD+GRPO training is said to improve Qwen3 research-plan quality, though the gains rest on LLM judgment without significance testing. Together the papers define a coherent agenda — getting more from smaller models with less data and clearer evaluation criteria — spanning Text-to-SQL semantic parsing and AI-for-science generation. The bibliometric snapshot is narrow: focus scores 100/100 (a single recorded field), collaboration 54/100 across 14 distinct coauthors, and reproducibility 55/100 averaged over one paper. Output, impact, and activity axes lack sufficient data and are unscored. No institution or country is recorded, so this profile is name-resolved from public bibliographic data.
Collaboration pattern
Collaboration scores 54/100, with 14 distinct coauthors in the snapshot. The frequent-coauthor list, after removing the researcher's own duplicated entry, centers on a recurring team — Weiming Lu, Yuhan Wang, Zhengxi Lu, Yuchen Yan, Kaitao Song, Wenqi Zhang, Jun Xiao, Yueting Zhuang, Haoyuan Ma, Hengwei Liu, Linjuan Wu — consistent with a stable group behind both 2026 preprints. Roles vary: middle author on LIMIT, last author on PaperGym, corresponding author on neither.
Research keywords
Highlights
796 curated samples reach 69.1% BIRD execution accuracy on Qwen3-8B
Rubric-as-reward training with 3.7% criterion leakage on PaperGym
One coherent data-centric post-training agenda across two 2026 preprints
The two reviewed papers share a thesis — smaller models, less data, clearer criteria: LIMIT cuts instruction-tuning data for Text-to-SQL, and PaperGym builds rubric-based rewards for research-plan generation. Both carry medium repro-readiness ratings, and the snapshot's single recorded field (NLP / Text-to-SQL) backs the concentration.
Timeline
- 2026PaperGym: Rubric-Centered Evolution for Research-Plan Generation appears on arXiv (the 2608 identifier in the catalog slug indicates August 2026); Yongliang Shen is the last author, not corresponding.
- 2026LIMIT: Less Is More for Instruction Tuning in Text-to-SQL appears on arXiv (the 2609 identifier in the slug indicates September 2026); middle-author position, not corresponding.
- 2026Evolve reviews both preprints; the reproducibility snapshot covers one paper with a mean score of 55, and 14 distinct coauthors are recorded over the coverage window.
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 is name-resolved from public bibliographic data; institution, country, and ORCID are missing in the source record, and the name is shared by multiple NLP researchers.
- Radar scores come from a small bibliometric snapshot: reproducibility averages only 1 paper, and the output, impact, and activity axes lack sufficient data and are left unscored.
- Coverage is limited to the two 2026 arXiv preprints reviewed by Evolve; peer-reviewed venue status, citation counts, and h-index are absent from the input data.
- Timeline ordering within 2026 is inferred from the arXiv identifiers embedded in the catalog slugs, and the frequent-coauthor list was deduplicated because it contains the researcher's own name.