Early-career researcher at the intersection of LLM post-training RL, inference serving, and LLM agents
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
6
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—
Not yet scored. Aggregate output measures are insufficient in the current OpenAlex record; the available h-index (6) and i10-index (2) are too thin to characterize productivity. This axis would measure total publication volume and its trajectory over time.
Impact—
Not yet scored. Citation-based impact aggregates (citations per year, field-normalized percentiles) are not yet available; h-index 6 and i10-index 2 indicate a very early citation record. This axis would measure influence on subsequent research via citations and downstream use.
Focus0/100
Scores 0/100, reflecting an evenly split field distribution: reinforcement learning for LLM post-training, ML systems for LLM inference and serving, and NLP/LLM agents each account for one third (0.33) of the record. The three 2026 cataloged papers map one-to-one onto these areas, indicating parallel rather than concentrated work; the OpenAlex topic list (topic modeling, parallel computing, NLP, information retrieval, oxides) widens the spread further.
Collaboration66/100
Scores 66/100. The record lists 26 distinct coauthors, with a stable core of twelve recurring collaborators including Yao Hu, Mu Chuan, Jiayan Fu, and Hang Xu. Notably, all three 2026 cataloged preprints place this researcher in a middle author position with no corresponding role, consistent with contributing to larger team efforts rather than solo-led projects.
Activity86/100
Scores 86/100, the strongest measured axis. OpenAlex indexes 14 works in 2026 — the highest single-year count in a record spanning 2013–2015 and 2022–2026 — and 18 works over the trailing three years. The 2026 cluster aligns with the three cataloged arXiv preprints, indicating a steeply accelerating current output cadence.
Reproducibility55/100
Scores 55/100, computed from Evolve's reproducibility assessments of three cataloged papers (reproPapers 3, mean 55); all three are rated medium readiness. Recurring limitations cited in review are single-run evaluations without error bars (PACT, Iris) and missing system baselines with tiny quality samples (RedKnot-MLA), though careful evidence labeling was commended.
Research summary
Zhaokai Luo (King University, US) is an early-career researcher whose OpenAlex field distribution splits evenly across three areas: reinforcement learning for LLM post-training, machine-learning systems for LLM inference and serving, and NLP/LLM agents. The three cataloged 2026 arXiv preprints map one-to-one onto these areas: PACT (axiom-grounded token-level credit assignment for actor-critic post-training), RedKnot-MLA (multi-head offline-online KV reuse for DeepSeek-V4 long-context serving), and Iris (open-source search agents). Activity is the strongest measured axis (86/100): OpenAlex indexes 14 works in 2026 and 18 works over the trailing three years, by far the densest cluster in the record. Collaboration scores 66/100, with 26 distinct coauthors and a recurring core of twelve, though all three cataloged papers place the researcher in a middle, non-corresponding author position. Reproducibility readiness is medium (mean 55/100 across three assessed papers), with Evolve reviews citing single-run evaluations and missing system baselines as recurring limitations. Output and impact axes remain unscored because aggregate citation data are thin (h-index 6, i10-index 2 in OpenAlex). The record shows a 2013–2015 cluster, a gap, then a 2022–2026 cluster, so early records may include homonyms; this profile is name-resolved from public bibliographic data.
Collaboration pattern
Collaboration is team-based: 26 distinct coauthors appear in the OpenAlex record, anchored by a recurring circle of twelve — Yao Hu, Mu Chuan, Jiayan Fu, Hang Xu, Yong Zhang, Dongyan Zhao, Yang Liu, Huayi Jin, Ruozhou He, Chenchen Hong, Mingxiao Ma, and Biao Zhang. All three 2026 cataloged preprints list the researcher as a middle author rather than corresponding author, consistent with contributing specialist expertise in RL post-training, serving systems, and agents within larger groups.
Research keywords
Highlights
Axioms that uniquely fix token-level credit in LLM post-training
Exact head-level KV reuse for DeepSeek-V4 long-context serving
Open-source search agents claimed at BrowseComp 82.2/88.6
Timeline
- 2026PACT preprint introduces axiom-grounded token-level credit assignment for actor-critic LLM post-training, with theory judged sound and strong empirical gains.
- 2026RedKnot-MLA preprint reports 2.02–3.84x TTFT gains from exact multi-head offline-online KV reuse for DeepSeek-V4 long-context serving.
- 2026Iris preprint claims state-of-the-art open-source search agents (BrowseComp 82.2/88.6) built on synthetic multi-hop data and an SFT-RL climbing recipe.
- 2026OpenAlex indexes 14 works in 2026 — the record's highest single-year count, within 18 works over the trailing three years — marking a steep activity peak.
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.
- All bibliometrics (h-index 6, i10-index 2, counts by year, topics, 26 distinct coauthors) come from a single OpenAlex name-resolved record; no ORCID was available to confirm identity.
- The record shows a 2013–2015 publication cluster, a gap through 2021, then a 2022–2026 cluster, and includes seemingly unrelated topics (e.g., electronic and structural properties of oxides), so early or peripheral records may reflect homonyms.
- Output and impact radar axes are unscored because aggregate citation and productivity data are insufficient; only h-index 6 and i10-index 2 are on file.
- Evolve's paper-level evidence covers three 2026 arXiv preprints only, none with verified peer-review status, and all rated medium reproducibility readiness.