Kuaishou researcher applying reinforcement learning across auto-bidding, LLM agents, and multi-agent LLM post-training
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
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—
No publication-volume data is available yet: this snapshot lacks total paper counts and yearly output records. This axis would measure papers published per year and their distribution across venues and authorship roles.
Impact—
Citation and influence metrics are not available in this snapshot, so no impact score can be computed. This axis would measure citation counts, aggregate citation-style indicators, and uptake of the reviewed papers by other researchers.
Focus0/100
The focus axis scores at the bottom of the range (0/100), and the raw field distribution explains why: the three reviewed papers split evenly across computational advertising/reinforcement learning, LLM agents, and reinforcement learning for LLMs, at about one-third (0.33) each. Within that spread, the lines still share reinforcement learning and credit assignment, so the dispersion runs across adjacent applied ML topics rather than unrelated disciplines.
Collaboration55/100
The collaboration axis scores 55/100, based on 15 distinct coauthors in the snapshot. The recurring pool is stable — Peng Jiang, Yewen Li, Shengtian Yang, Lei Feng, Yitian Li, Pengfei Lv, Longtao Zheng and Kun Gai appear repeatedly — and all three reviewed papers are multi-author efforts, with Qingpeng Cai in the last position on OneBid and middle positions on SRPO and AgentBrew.
Activity—
No year-by-year productivity data is available yet, so recent research cadence cannot be scored. This axis would measure publication frequency over time and continuity of research; the snapshot currently shows only three papers, all dated 2026.
Reproducibility43/100
Reproducibility readiness averages 43/100 across the three reviewed papers. OneBid rates low — headline production gains are proprietary and offline checks use critic-based proxies; SRPO and AgentBrew rate medium, with public model baselines but gains that are either sub-noise against non-rerun baselines or drawn from small evaluations lacking variance reporting.
Research summary
Qingpeng Cai is a researcher at Kuaishou (China) whose reviewed work spans three fields in near-equal measure: computational advertising with reinforcement learning, LLM agents, and reinforcement learning for LLMs, each holding about one-third of the bibliometric field distribution. The Evolve catalog includes three 2026 arXiv preprints. As last author on OneBid, he is positioned on a unified auto-bidding foundation model for diverse oCPX advertising scenarios whose production A/B tests report a +2.2% ADVV gain, although Evolve notes the evidence is proprietary and offline checks rely on critic-based proxies. SRPO reformulates multi-agent LLM post-training through set-valued actions, applying one √K-normalized clip per transition; Evolve calls the formulation elegant and practical while measuring its math gains as sub-noise against non-rerun baselines. AgentBrew trains tool-use agents offline from raw real-world trajectories using hindsight relabeling and PMI action credits, lifting Qwen3-32B by +8.7 average accuracy across three MCP applications, on small evaluations that lack variance reporting. With 15 distinct coauthors, recurring collaborators include Peng Jiang, Yewen Li, Shengtian Yang, Lei Feng, Yitian Li, Pengfei Lv, Longtao Zheng and Kun Gai. Average reproducibility readiness across the three reviewed papers is 43 out of 100.
Collaboration pattern
Collaboration is industry-anchored: within 15 distinct coauthors, Peng Jiang, Yewen Li, Shengtian Yang, Lei Feng, Yitian Li, Pengfei Lv, Xialong Liu, Zhiyi Lyu, Longtao Zheng, Lang Feng and Kun Gai recur. All three reviewed papers are multi-author efforts, with Qingpeng Cai last-authoring OneBid and holding middle positions on SRPO and AgentBrew — a placement pattern consistent with a senior individual contributor or team-lead role rather than sole first-author work.
Research keywords
Highlights
OneBid: unified auto-bidding foundation model for oCPX advertising
SRPO: set-valued actions for multi-agent LLM policy optimization
AgentBrew: offline tool-use agent learning from raw trajectories
Timeline
- 2026OneBid preprint claims the first auto-bidding foundation model for diverse oCPX advertising scenarios, reporting +2.2% ADVV in production A/B tests.
- 2026SRPO introduces setwise relative policy optimization for multi-agent LLMs, with a single √K-normalized clip per transition.
- 2026AgentBrew reports offline tool-use agent learning from raw real-world trajectories, gaining +8.7 average accuracy on Qwen3-32B across three MCP apps.
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.
- Bibliometrics and radar scores derive from an OpenAlex-based bibliometric snapshot only; the output, impact and activity axes lack underlying data and are not scored.
- Profile scope is limited to the three 2026 arXiv preprints reviewed by Evolve; earlier publications, citation records and peer-reviewed venues are not covered.
- Identity is name-resolved via ORCID 0000-0001-6451-9299 from public bibliographic data; institution and country are as listed in the snapshot.
- The reproducibility score (43) reflects Evolve readiness ratings on these three papers only and may not generalize to the wider body of work.