Zhejiang University researcher advancing data-efficient instruction tuning, agentic RL, and LLM-based evaluation
4
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
No output score yet: the OpenAlex snapshot lacks usable publication-volume and citation metrics (h-index and i10-index recorded as 0), so this axis, which would measure total and per-year scholarly output, cannot be scored.
Impact—
Impact data is insufficient: citation counts are absent from the record, so influence on follow-up work cannot be estimated. This axis would measure citations, citations per paper, and uptake of the reviewed methods by other teams.
Focus0/100
The focus score is 0, and the underlying field distribution explains why: effort splits evenly (~33% each) across machine translation evaluation, LLM post-training/agentic RL, and Text-to-SQL semantic parsing. Topic counts reinforce this breadth — natural language processing techniques (8), topic modeling (7), multimodal ML applications (6), reinforcement learning in robotics (5), and advanced neural network applications (3) — indicating an actively broad rather than concentrated 2026 portfolio.
Collaboration68/100
Collaboration scores 68, based on about 30 distinct coauthors in the recent three-year window. The frequent-coauthor list centers on Wenqi Zhang, Yongliang Shen, Yizhou Liu, Yueting Zhuang, and Jun Xiao, alongside Kaitao Song, Yuchen Yan, and others, suggesting a stable Zhejiang University cluster. Across the four catalogued 2026 papers, Weiming Lu appears in middle author positions three times and once as last author, consistent with a busy, well-connected collaborator within a larger group.
Activity89/100
Activity is the strongest axis at 89. The record lists 20 works in the recent three-year window, with the 2026 count alone at 20 — effectively all recent output concentrated in the current year. The Evolve catalog independently confirms the pace, holding four 2026 arXiv papers spanning Text-to-SQL tuning, agentic RL distillation, MT evaluation, and research-plan generation, all within a single year of preprint activity.
Reproducibility43/100
Reproducibility scores 43 (reproMean) across 3 papers assessed. Readiness in the catalog is medium for LIMIT, RetireOPD, and PaperGym, and low for the CuRIL translation-evaluation work. Recurring review flags — single-run RL experiments without variance, missing significance tests, and limited public data or code — suggest promising pipelines whose artifacts are not yet fully releasable.
Research summary
Weiming Lu is a researcher at Zhejiang University whose recent work sits at the intersection of natural language processing and machine learning; this profile is name-resolved from OpenAlex bibliographic data. The record shows 20 works in the recent three-year window, all tallied in 2026, alongside about 30 distinct coauthors, yielding an activity score of 89 and a collaboration score of 68. The Evolve catalog holds four 2026 arXiv papers spanning three roughly equal fields (33% each): machine translation evaluation, LLM post-training/agentic reinforcement learning, and Text-to-SQL semantic parsing. Recurring themes include data-efficient instruction tuning (LIMIT reaches 69.1% BIRD execution accuracy on Qwen3-8B with 796 curated samples), on-policy distillation for agentic RL (RetireOPD reports +12–19 points over GRPO on ALFWorld/WebShop), culturally grounded evaluation of social-media translation (CuRIL), and rubric-centered training for research-plan generation (PaperGym, 3.7% criterion leakage). Authorship across the catalogued papers — middle three times, last once, none corresponding — fits a hands-on contributor within a larger lab cluster that includes Wenqi Zhang, Yongliang Shen, Yueting Zhuang, and Jun Xiao. Citation-impact metrics are unavailable (h-index and i10-index recorded as 0, indicating missing coverage rather than measured zeros), so this profile emphasizes current output and methodological range over accumulated influence.
Collaboration pattern
About 30 distinct coauthors appear in the recent three-year window. The recurring group is anchored at Zhejiang University — Wenqi Zhang, Yongliang Shen, Yizhou Liu, Yueting Zhuang, Jun Xiao, Kaitao Song, Yuchen Yan — with additional partners including Yuhan Wang, Zhengxi Lu, Yiwen Qiu, and Linjuan Wu. Across the four catalogued 2026 papers, Weiming Lu holds middle authorship three times and last authorship once, with no corresponding-author role, consistent with a contributing coauthor within a large, active lab.
Research keywords
Highlights
796 curated samples reach 69.1% BIRD execution accuracy on Qwen3-8B
+12–19 points over GRPO via self-retiring on-policy distillation
Metrics and LLM judges fail on social-media machine translation
Rubric-centered training improves research-plan quality with 3.7% leakage
Timeline
- 2026Publishes LIMIT: less-is-more instruction tuning for Text-to-SQL, with 796 curated samples reaching 69.1% BIRD execution accuracy on Qwen3-8B (last author)
- 2026Releases RetireOPD, self-retiring on-policy distillation for agentic RL, reporting +12–19 points over GRPO on ALFWorld/WebShop
- 2026Proposes CuRIL, internalizing cultural reasoning for social-media translation evaluation, showing metrics and LLM judges fail without cultural grounding
- 2026Introduces PaperGym, rubric-centered evolution for research-plan generation, pairing 3.7% criterion-leakage rubric data with OPSD+GRPO training
Papers in the Evolve catalog
LIMIT: Less Is More for Instruction Tuning in Text-to-SQL
RetireOPD: Self-Retiring On-Policy Distillation for Agentic Reinforcement Learning
When Metrics Reward the Worst Translations: Internalizing Cultural Reasoning for Social Media Translation Evaluation
PaperGym: Rubric-Centered Evolution for Research-Plan Generation
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 signals come from OpenAlex only; citation-based metrics (h-index, i10-index) are recorded as 0, reflecting missing coverage rather than measured zeros, so the output and impact axes are omitted.
- All four catalogued works are 2026 arXiv preprints; peer-reviewed venue status, and any pre-2026 publication record, is not captured in the provided data.
- The country code in the source record ("IL") is inconsistent with the listed institution (Zhejiang University); the profile relies on the institution string.
- Identity is name-resolved from public bibliographic data (no ORCID available); "Weiming Lu" is a fairly common Chinese name, so attribution carries some ambiguity.