Efficient LLM inference researcher: low-bit quantized reasoning and MoE expert pruning; last author on 2026 arXiv.
2
papers here
2
reviewed by Evolve
0
first-author
0
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—
No output metrics are available in this snapshot. This axis would measure publication volume and authorship roles over time; with no per-year counts to ground it, the dimension is left unscored rather than estimated.
Impact—
Impact data such as citation counts, h-index, or downstream adoption is not provided. This axis would measure the citation-based influence of the two cataloged papers; it is withheld rather than guessed.
Focus0/100
The focus score is 0, yet the raw field distribution splits 0.5/0.5 between two nearly identical labels — machine learning (efficient LLM inference) and machine learning — efficient LLM inference and model compression — and both cataloged papers sit in this one topic. The portfolio is topically tight; the zero score appears to reflect a very thin publication record (two papers) rather than scattered research interests.
Collaboration60/100
Collaboration scores 60, supported by 19 distinct coauthors in the record. Recurring names include Peisong Wang, Yuanteng Chen, and Yuantian Shao (each appearing twice in the frequent-coauthor list), alongside Chuangyi Li, Jing Liu, Qiwei Lai, Chen Tianqi, Nanxin Zeng, and Zhilei Liu. Both cataloged papers place Jian Cheng in the last author position without corresponding status, consistent with a senior, supervision-oriented role in a stable working group.
Activity—
Recent-activity signals, such as publications per year or preprint cadence, are absent from the input. This axis would measure ongoing research output over time; only two 2026 papers are recorded here, so no score is given.
Reproducibility55/100
Reproducibility scores 55, a reproMean computed over 2 papers, both rated medium readiness. Evolve's review of the distillation study notes 'well-controlled but single-run evidence and partial ablations', while the MoE pruning study is judged 'convincing, solid work' — careful empirical design whose full independent replication remains to be demonstrated.
Research summary
Jian Cheng is a machine-learning researcher working on efficient large language model inference. The cataloged record centers on two 2026 arXiv papers, both with Jian Cheng as last author (non-corresponding). The first, 'Train Where the Quantized Model Goes', combines quantization-aware distillation (QAD) with on-policy distillation to restore reasoning in sub-3-bit quantized models, lifting MATH retention from 35% to 70%; Evolve's review calls the evidence well-controlled but notes single-run results and partial ablations. The second, an empirical study of dynamic expert pruning in fine-grained MoE LLMs, shows that uniformly keeping 2/3 of the chosen experts preserves 98.8% of MoE performance, with dynamic pruning paying off only under aggressive budgets; Evolve rates it convincing, solid work. Together they define a tight agenda spanning model compression (extreme low-bit quantization, distillation) and architecture-level efficiency (mixture-of-experts pruning), evaluated with a controlled, iso-budget empirical style. Radar dimensions from the bibliometric record give collaboration 60 (19 distinct coauthors, with Peisong Wang, Yuanteng Chen, and Yuantian Shao recurring), reproducibility 55 (mean across the two papers, both medium readiness), and focus 0, which sits against a field distribution split 0.5/0.5 between two labels of essentially the same topic. Output, impact, and activity lack underlying metrics and are left unscored.
Collaboration pattern
Jian Cheng appears as last (non-corresponding) author on both cataloged papers, consistent with a senior supervisory role. The record shows 19 distinct coauthors; the frequent list centers on Peisong Wang, Yuanteng Chen, and Yuantian Shao (each repeated), with Chuangyi Li, Jing Liu, Qiwei Lai, Chen Tianqi, Nanxin Zeng, and Zhilei Liu also recurring — a stable, moderately sized network dedicated to efficient LLM inference.
Research keywords
Highlights
On-policy distillation restores sub-3-bit reasoning
2/3 of chosen experts keep 98.8% of MoE performance
A tight empirical agenda in efficient LLM inference
Across both 2026 papers — extreme low-bit quantization with distillation, and expert pruning in fine-grained MoE — a consistent style emerges: controlled experiments, explicit budget comparisons, and last-author supervision of the studies.
Timeline
- 2026Publishes (last author, arXiv) 'Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning'.
- 2026Publishes (last author, arXiv) the empirical study 'You Only Need 2/3 of the Chosen Experts' on dynamic expert pruning in fine-grained MoE LLMs.
- 2026Both cataloged papers land in efficient LLM inference and model compression, forming a 0.5/0.5 field distribution in the bibliometric record.
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.
- Bibliometric scores are computed from a sparse public bibliographic snapshot; only two reviewed papers feed this profile.
- No institution, country, or ORCID was provided; 'Jian Cheng' is a common name, so records are name-resolved from public bibliographic data and homonyms may be mixed.
- Output, impact, and activity axes lack underlying metrics (no citation counts or per-year publication volumes), so they are left unscored.
- The reproducibility mean (55) rests on only two papers, both rated medium readiness; peer-review and venue acceptance beyond arXiv are unknown.