Peking University researcher in LLM agents, prompt optimization, and video-generation reasoning evaluation
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
30
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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—
This axis is not scored: the underlying output data is insufficient in the current snapshot. It would measure total publication volume, venue tiers, and publication-type mix over the career, beyond the raw annual work counts already visible in the record.
Impact—
Impact is not scored because citation-based evidence is insufficient in this snapshot. The axis would measure citation counts, citation trajectory relative to field norms, and uptake of the researcher's methods in follow-up work.
Focus0/100
The focus score is 0/100, reflecting high topical dispersion: the record splits evenly across three fields — machine learning (LLM agents), artificial intelligence (LLM agent systems), and computer vision (video generation evaluation) at one third each — while five OpenAlex topic clusters span recommender systems (166), information retrieval (150), bandit algorithms (110), GANs (55), and multimodal learning (49). The three reviewed papers likewise cover three distinct fields.
Collaboration66/100
Collaboration scores 66/100, based on 27 distinct coauthors recorded over the last three years. Twelve frequent collaborators appear in the record, including Zhiyi Lyu, Yewen Li, Longtao Zheng, Shengtian Yang, Lang Feng, Lei Feng, Peng Jiang, Qingpeng Cai, Bo An, Zibo Zhao, Jijun Shi, and Mo Zhou. On the reviewed papers Gai holds middle and last authorship, with no corresponding-author role flagged.
Activity100/100
Activity scores 100/100, the maximum. OpenAlex counts 202 works in the last three years with an accelerating curve: 23 works in 2023, 37 in 2024, 64 in 2025, and 101 in 2026 — the highest annual figure in a record reaching back to a single work in 2008, with a publication gap across 2014–2016.
Reproducibility43/100
Reproducibility averages 43/100 across the three Evolve-reviewed papers. Readiness is rated medium for AgentBrew and for the think-with-video benchmark, and low for RobustSGPO. Recurring concerns include a judge-circular, statistically thin 380-sample evaluation, small evaluations without variance reporting, and single-workflow evidence without significance testing.
Research summary
Kun Gai is a researcher at Peking University whose OpenAlex-resolved record shows an h-index of 30 and an i10-index of 57, with sharply accelerating output: 23 works in 2023, 37 in 2024, 64 in 2025, and 101 in 2026, totalling 202 works over the last three years. The long-run topic profile is anchored in recommender systems and techniques (166), information retrieval and search behavior (150), and bandit algorithms (110), with smaller clusters in GANs and image synthesis (55) and multimodal machine learning (49); the record reaches back to 2008. The three papers Evolve has reviewed — all 2026 arXiv releases — indicate a current center of gravity in LLM agents and generative-model evaluation: AgentBrew on offline tool-use agent learning from raw real-world trajectories (middle author), RobustSGPO on search-space control for agent harness evolution (last author), and a nine-dimension think-with-video reasoning benchmark with a text-reward prompt rewriter (middle author). Evolve's reproducibility assessment averages 43/100 across these three papers, citing a judge-circular 380-sample evaluation, small evaluations without variance reporting, and single-workflow evidence. Recent collaboration spans 27 distinct coauthors.
Collaboration pattern
OpenAlex records 27 distinct coauthors over the last three years, yielding a collaboration score of 66/100. Twelve frequent collaborators appear, including Zhiyi Lyu, Yewen Li, Longtao Zheng, Shengtian Yang, Lang Feng, Lei Feng, Peng Jiang, Qingpeng Cai, Bo An, Zibo Zhao, Jijun Shi, and Mo Zhou. Across the three reviewed 2026 papers Gai appears in middle and last authorship positions with no corresponding-author role flagged, consistent with substantial but distributed involvement.
Research keywords
Highlights
AgentBrew: offline tool-use agent learning from raw real-world trajectories
RobustSGPO: search-space control for agent harness evolution
Benchmarking and improving think-with-video reasoning for video generative models
Timeline
- 2026AgentBrew — offline tool-use agent learning from raw real-world trajectories (middle author, arXiv preprint).
- 2026RobustSGPO — search-space control for agent harness evolution (last author, arXiv preprint).
- 2026Think-with-video reasoning benchmark and prompt rewriter for video generative models (middle author, arXiv preprint).
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 indicators (h-index, i10 index, yearly counts, topics, co-author counts) come from OpenAlex only; the 'output' and 'impact' dimension scores are unavailable in this snapshot.
- Qualitative judgments rest on three Evolve-reviewed papers, all 2026 arXiv preprints; the reproducibility mean of 43 covers only these three works.
- The profile is name-resolved from public bibliographic data (OpenAlex identity, ORCID 0000-0002-3636-3618); no biographical verification beyond these records.
- Historical topic clusters (recommender systems, information retrieval, bandits) and the current LLM-agent focus coexist; the even three-way field split makes field attribution uncertain.