Peisong Wang

Chinese Academy of Sciences

Unclaimed · AI-structured profileOpenAlex

Chinese Academy of Sciences researcher on efficient LLM inference: low-bit quantization recovery and MoE pruning

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-08) 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 scored: publication-volume data is insufficient in this snapshot. This axis would measure total research output — number of papers, venues, and authorship roles over time — none of which is available beyond the two reviewed 2026 arXiv preprints.

Impact—

Not scored: citation-based impact data is unavailable. This axis would measure citations, h-index-style aggregates, and influence on follow-up work; no citation counts or venue-prestige records are present in the input, so no estimate is offered.

Focus0/100

The recorded focus score is 0/100. The raw field distribution is maximally concentrated: mass splits 0.5/0.5 across two nearly identical labels — "machine learning (efficient LLM inference)" and "machine learning — efficient LLM inference and model compression" — and both 2026 catalog papers (quantized-model distillation; MoE expert pruning) belong to that same efficient-inference lane. Within the observed window this reads as a single-subfield specialist rather than a generalist.

Collaboration60/100

Collaboration scores 60/100, backed by 19 distinct coauthors in the raw record. The recurring-collaborator list clusters around a stable core — Jian Cheng, Jing Liu, Chuangyi Li, Yuanteng Chen, Yuantian Shao, Qiwei Lai, Nanxin Zeng, Zhilei Liu, and Chen Tianqi — suggesting repeat partnerships rather than one-off ties. Both reviewed 2026 papers are middle-author, non-corresponding contributions to multi-author teams, consistent with a moderate, team-embedded collaboration pattern.

Activity—

Not scored: yearly publication-cadence data is insufficient. This axis would measure recent output rate and research tempo — papers per year and continuity of activity — but only two 2026 items are on record, which cannot establish a trend.

Reproducibility55/100

Reproducibility averages 55/100 across the 2 papers Evolve has scored (reproPapers = 2, reproMean = 55); both are rated medium reproducibility readiness. Evolve's verdicts add nuance: the distillation paper reports strong MATH-retention gains (35%→70%) but rests on single-run evidence with partial ablations, while the MoE pruning study is called "convincing, solid work" on the strength of iso-budget evaluation. Promising, but not yet independently verified.

Research summary

Peisong Wang is a researcher at the Chinese Academy of Sciences working on efficient large language model (LLM) inference and model compression. The bibliometric field distribution splits evenly (0.5/0.5) between "machine learning (efficient LLM inference)" and "machine learning — efficient LLM inference and model compression", and both reviewed papers sit in that same narrow lane, consistent with the recorded focus score of 0/100 — a highly specialized profile. Two 2026 arXiv preprints define the current catalog: one shows that on-policy distillation after quantization-aware distillation restores sub-3-bit reasoning, lifting MATH retention from 35% to 70% (Evolve notes well-controlled but single-run evidence with partial ablations); the other, an empirical study of dynamic expert pruning in fine-grained MoE LLMs, finds that uniformly retaining 2/3 of chosen experts preserves 98.8% of performance, with dynamic pruning paying off only under aggressive budgets. Reproducibility averages 55/100 across the two scored papers, both rated medium readiness; collaboration scores 60/100 with 19 distinct coauthors and a stable recurring core including Jian Cheng, Jing Liu, Chuangyi Li, Yuanteng Chen, and Yuantian Shao. Output, impact, and activity axes lack sufficient data and are not scored. On both papers Wang is a middle-position, non-corresponding author, so independent research leadership cannot yet be assessed from this record.

Collaboration pattern

Collaboration scores 60/100 with 19 distinct coauthors on record. Work flows through a recurring core — Jian Cheng, Jing Liu, Chuangyi Li, Yuanteng Chen, Yuantian Shao, Qiwei Lai, Nanxin Zeng, Zhilei Liu, and Chen Tianqi — indicating stable repeat partnerships rather than ad hoc teaming. On both reviewed 2026 papers Wang contributes as a middle-position, non-corresponding author within multi-author arXiv teams, suggesting a team-embedded contributor rather than a solo or corresponding-author lead so far.

Research keywords

quantization-aware distillationon-policy distillationextreme low-bit quantizationexposure biasLLM reasoningreinforcement learning with verifiersmixture-of-expertsdynamic expert pruningLLM inference efficiencyfine-grained MoEexpert selection redundancyiso-budget evaluation

Highlights

On-policy distillation restores reasoning after extreme quantization
In the 2026 arXiv paper "Train Where the Quantized Model Goes", on-policy distillation applied after quantization-aware distillation recovers sub-3-bit reasoning, lifting MATH retention from 35% to 70%; Evolve calls the study well-controlled but notes single-run evidence and partial ablations.
Shows 2/3 of chosen experts suffice in fine-grained MoE LLMs
The 2026 empirical study of dynamic expert pruning finds that uniformly retaining 2/3 of the chosen experts preserves 98.8% of MoE performance, while dynamic pruning pays off only under aggressive budgets — a result Evolve rates convincing and solid, grounded in iso-budget evaluation.

Timeline

2026Publishes on arXiv "Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning" — middle-position, non-corresponding author; Evolve rates it medium reproducibility readiness.
2026Publishes on arXiv the empirical study "You Only Need 2/3 of the Chosen Experts" on dynamic expert pruning in fine-grained MoE LLMs — again as a middle-position, non-corresponding author.
2026Both preprints enter the Evolve review catalog; the two reproducibility scores yield a 55/100 mean, and the coauthor record reaches 19 distinct names.

Papers in the Evolve catalog

Frequent co-authors

Yuanteng Chen×2 Yuantian Shao×2 Chuangyi Li×2 Jing Liu×2 Jian Cheng×2 Yuanteng Chen×1 Qiwei Lai×1 Chen Tianqi×1 Peisong Wang×1 Yuantian Shao×1 Nanxin Zeng×1 Zhilei Liu×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 computed from public bibliometric data (OpenAlex-style) only; no ORCID is provided, and "Peisong Wang" is a common Chinese name, so identity resolution carries some uncertainty.
Analysis scope covers only the two Evolve-reviewed 2026 arXiv preprints; publication volume, citations, h-index, and venue history are absent, leaving output, impact, and activity unscored.
Both papers are arXiv preprints with unknown peer-review status, and Wang is a middle-position, non-corresponding author on each, limiting inference about research leadership.
The raw frequent-coauthor list contained duplicates and the subject's own name; after removal, 9 recurring collaborators remain, all reported as given without affiliation claims.