Peng Jiang

Kuaishou (China)

Unclaimed · AI-structured profileOpenAlex

Kuaishou researcher applying reinforcement learning to auto-bidding, online advertising, and LLM tool-use agents

papers here

reviewed by Evolve

first-author

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—

Not yet scored: total output volume, venue mix, and cumulative publication counts are unavailable — the current data covers only three 2026 arXiv papers in the Evolve catalog, with no verified full publication list.

Impact—

Not yet scored: citation counts, h-index, and downstream adoption signals are absent from the input data. This axis would measure scholarly influence and real-world uptake; with evidence confined to recent preprints, no influence estimate is possible.

Focus0/100

Focus scores 0 on the 0-100 scale, reflecting a field distribution split evenly at about 33% each across computational advertising/reinforcement learning, machine learning (LLM agents), and recommender systems/online advertising. All three sit within applied machine learning for advertising and agents, but with only three 2026 papers no single field yet dominates, so the record reads as broad applied ML rather than a concentrated niche.

Collaboration60/100

Collaboration scores 60, computed from 19 distinct coauthors across the three 2026 papers. Recurring names include Qingpeng Cai, Yewen Li, Yitian Li, Pengfei Lv, Xialong Liu, and Kun Gai, pointing to a stable Kuaishou industrial team. Jiang sits mid-list on OneBid and AgentBrew and is last author on DMRL, with no corresponding-author role recorded, consistent with team-based industrial research.

Activity—

Not yet scored: yearly publishing tempo cannot be computed because the snapshot contains papers from a single year (2026). This axis would measure output cadence and sustained research engagement across multiple years.

Reproducibility32/100

Reproducibility readiness averages 32 across three papers: AgentBrew rates medium, while OneBid and DMRL rate low. Evidence rests on proprietary Kuaishou serving platforms; OneBid's offline checks are critic-based proxies, DMRL reports a +0.052% LTV effect without confidence intervals or significance tests, and AgentBrew's small evaluations lack variance reporting, all of which limit independent replication.

Research summary

Peng Jiang is an applied machine-learning researcher at Kuaishou (China) whose 2026 work spans computational advertising, recommender systems, and LLM agents. In the Evolve catalog Jiang appears on three arXiv papers, all from 2026. OneBid proposes a unified auto-bidding foundation model for diverse oCPX advertising scenarios, combining Decision Transformer architectures, Mixture-of-Experts, and offline reinforcement learning; Evolve notes it claims to be the first auto-bidding foundation model, with production A/B gains of +2.2% ADVV, though the evidence is proprietary and offline checks rely on critic-based proxies. AgentBrew addresses offline tool-use agent learning from raw real-world trajectories, using hindsight relabeling and PMI-based action credits to lift Qwen3-32B by +8.7 average accuracy across three MCP applications, with medium reproducibility readiness but no variance reporting on small evaluations. DMRL, where Jiang is last author, recasts LLM skill-document optimization in advertising recommendation as reinforcement learning via DRPO and LRP, reporting a deployed +0.052% LTV gain at Kuaishou. Collaboration data records 19 distinct coauthors, with a reproducibility mean of 32 across three papers, reflecting reliance on private platform evidence. Key quantitative claims are quoted from the source papers and not independently verified.

Collaboration pattern

Jiang researches within a stable Kuaishou applied-ML cluster: 19 distinct coauthors across three 2026 papers, with Qingpeng Cai, Yewen Li, Yitian Li, Pengfei Lv, Xialong Liu, Kun Gai, Lang Feng, and Shengtian Yang recurring. Authorship is mid-list on OneBid and AgentBrew and last on DMRL, with no corresponding-author role recorded — a pattern consistent with industrial team science on advertising systems and LLM agents rather than a solo academic line.

Research keywords

auto-biddingoCPX advertisingfoundation modelDecision TransformerMixture-of-Expertsoffline reinforcement learningLLM agentstool useModel Context Protocol (MCP)hindsight relabelingcredit assignmentdelayed feedback

Highlights

OneBid: unified auto-bidding foundation model for diverse oCPX advertising
Positioned as the first auto-bidding foundation model, OneBid combines Decision Transformer, Mixture-of-Experts, and offline reinforcement learning across diverse oCPX scenarios; production A/B testing shows +2.2% ADVV, though evidence is proprietary and offline checks rely on critic-based proxies (reproducibility: low).
AgentBrew: offline tool-use agent learning from raw real-world trajectories
Hindsight relabeling plus PMI-based action credits turn raw real-world trajectories into offline training signal, lifting Qwen3-32B by +8.7 average accuracy across three MCP applications; the most reproducible of Jiang's catalog papers (medium), though small evaluations lack variance reporting.
DMRL: skill-document tuning in ad recommendation recast as reinforcement learning
As last author, Jiang contributed to DMRL, which optimizes LLM skill documents via DRPO and LRP under delayed feedback; it is deployed on Kuaishou advertising with a reported +0.052% LTV gain — a tiny effect reported without confidence intervals on a private platform (reproducibility: low).

Timeline

2026Co-authored OneBid, a unified auto-bidding foundation model for diverse oCPX advertising scenarios, reporting +2.2% ADVV in production A/B tests.
2026Co-authored AgentBrew on offline tool-use agent learning from raw real-world trajectories, lifting Qwen3-32B by +8.7 average accuracy across three MCP applications.
2026Last-author DMRL recast skill-document optimization in advertising recommendation as reinforcement learning, deployed at Kuaishou with a reported +0.052% LTV gain.

Papers in the Evolve catalog

Frequent co-authors

Qingpeng Cai×2 Yewen Li×1 Yitian Li×1 Pengfei Lv×1 Xialong Liu×1 Zhiyi Lyu×1 Yewen Li×1 Longtao Zheng×1 Shengtian Yang×1 Lang Feng×1 Lei Feng×1 Kun Gai×1

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 derived from OpenAlex-based public bibliometric data; no citation counts, h-index, or venue-level metrics are included in this snapshot.
Coverage is limited to three 2026 arXiv papers in the Evolve catalog; earlier or venue-published output is not represented, which is why output, impact, and activity axes are unscored.
'Peng Jiang' is a common name and no ORCID is linked; this profile is name-resolved from public bibliographic data and identity consolidation should be verified before citing.
Headline results (+2.2% ADVV, +8.7 Acc avg, +0.052% LTV) are quoted from the papers as reviewed and come from proprietary Kuaishou platforms, without independent verification.