Microsoft researcher bridging LLM post-training (MoE RL exploration) and LLM inference systems (AI-driven architecting)
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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—
Output metrics are unavailable: beyond the two cataloged 2026 arXiv preprints, OpenAlex provides no aggregated publication-volume data for this author. This axis would measure total works, authorship roles, and venue spread.
Impact—
Impact data is insufficient: the record lists h-index and i10-index as 0 with no citation aggregation, so influence cannot be assessed. This axis would measure citations, citing works, and uptake of the MoE-RL and inference-systems papers.
Focus0/100
The focus score is 0, consistent with an even 50/50 fieldDist split between machine learning (LLM post-training) and computer systems (MLSys / LLM inference). Topic assignments are likewise dispersed — Natural Language Processing Techniques (2 works), plus one each in Topic Modeling, Advanced Neural Network Applications, Parallel Computing and Optimization Techniques, and Reinforcement Learning in Robotics — indicating two parallel research lines rather than a single concentrated agenda.
Collaboration44/100
Collaboration scores 44. The cataloged works involve 8 distinct coauthors, and Evolve lists 8 frequent collaborators including Hongyi He, Zhenghao Lin, Xiao Liu, Yan Lu, Yeyun Gong, Ziyue Yang, Yuting Jiang, and Lei Qu. Authorship positions are middle on the MoE-RL paper and last on RoofLang, with neither flagged as corresponding — consistent with contributing to team projects at Microsoft rather than solo-led efforts.
Activity32/100
Activity scores 32. countsByYear contains a single entry — 2026 with 2 works — and recent3yWorks is 2, so the observable record is two arXiv preprints posted in 2026. The snapshot holds no earlier-year publication data to establish a longer cadence, so this score reflects a short observable window rather than a measured decline.
Reproducibility55/100
Reproducibility scores 55 (reproMean over 2 reproPapers). Evolve rates both papers as medium reproReadiness: ESRL reports gains across three MoE backbones but from single runs without variance or significance tests, while RoofLang rests on uncalibrated discrete-event simulation and omits a promised grid-search comparison. The mid-range score reflects clearly described methods offset by unverified quantitative claims.
Research summary
Peng Cheng is a researcher at Microsoft (United States) whose cataloged work spans two complementary areas: LLM post-training via reinforcement learning and systems for LLM inference. OpenAlex attributes two recent works, both 2026 arXiv preprints, with 8 distinct coauthors and an even 50/50 fieldDist split between machine learning (LLM post-training) and computer systems (MLSys / LLM inference). On the post-training side, ESRL adds expert-routing exploration to MoE reinforcement learning built on GRPO; Evolve reports gains across three MoE backbones but flags single-run results without variance or significance testing. On the systems side, RoofLang pairs a domain-specific language with a roofline-based discrete-event simulator so AI agents can architect LLM inference systems; the review notes the simulations are uncalibrated and a promised grid-search comparison is absent. Radar dimensions place collaboration at 44 and reproducibility at 55 (mean over 2 papers), with activity at 32 — output concentrated in 2026 — and focus at 0, reflecting the even two-field split. Citation-based output and impact metrics are unavailable (h-index and i10-index recorded as 0), making this an early-stage, systems-and-algorithms profile.
Collaboration pattern
Peng Cheng's cataloged work is team-based: the two 2026 arXiv papers involve 8 distinct coauthors, with 8 frequent collaborators listed — Hongyi He, Zhenghao Lin, Xiao Liu, Yan Lu, Yeyun Gong, Ziyue Yang, Yuting Jiang, and Lei Qu. Neither paper credits Cheng as first or corresponding author (middle position on the MoE-RL paper, last on RoofLang), suggesting a contributor role within Microsoft-affiliated groups spanning both LLM post-training and inference systems.
Research keywords
Highlights
ESRL: Expert-Space Exploration for MoE Reinforcement Learning
RoofLang: A DSL and Simulator for AI-Driven LLM Inference Architecting
Timeline
- 2026Co-authored ESRL on expert-space exploration in MoE reinforcement learning (middle author, arXiv preprint; field: LLM post-training).
- 2026Senior-position (last author) contribution to RoofLang, a domain-specific language and discrete-event simulator for AI-driven architecting of LLM inference systems.
- 2026Both cataloged works appear in 2026: countsByYear records 2 works, split evenly between ML post-training and ML systems.
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 come from OpenAlex; h-index and i10-index are recorded as 0, which likely reflects thin author-disambiguation coverage rather than measured impact.
- "Peng Cheng" is a common name; this profile is name-resolved from public bibliographic data and may merge or omit records.
- Both reviewed works are 2026 arXiv preprints; no peer-reviewed venue, citation, or funding data is available.
- The reproducibility score averages only 2 papers (reproMean 55) and the output/impact axes are marked missing, so cross-author comparisons should stay preliminary.