Efficient LLM inference researcher at the Chinese Academy of Sciences: low-bit distillation and MoE expert pruning.
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 publication-count or output metrics are available yet, so this axis is unscored. It would measure the volume and consistency of Chuangyi Li's publications over time, including preprints and peer-reviewed papers.
Impact—
Citation and influence data are unavailable, so impact is unscored. This axis would measure citations, downstream adoption, and community uptake of the quantization-distillation and MoE-pruning findings.
Focus0/100
The radar reports a focus score of 0, while the raw field distribution splits evenly (0.5/0.5) between two labels: 'machine learning (efficient LLM inference)' and 'machine learning — efficient LLM inference and model compression'. Because both labels describe essentially the same topic, this score should be read cautiously and may reflect label fragmentation rather than genuine breadth; the cataloged work is in practice tightly concentrated on efficient LLM inference.
Collaboration60/100
The collaboration score is 60, based on 19 distinct coauthors in the bibliometric record. Recurring collaborators include Yuanteng Chen, Peisong Wang, Yuantian Shao, Jing Liu, Jian Cheng, Qiwei Lai, Chen Tianqi, Nanxin Zeng and Zhilei Liu. On both cataloged papers Li is a middle author without corresponding status, consistent with contributing within a team rather than leading it.
Activity—
Recent-activity data are insufficient, so this axis is unscored. It would measure publishing cadence and ongoing research momentum, which cannot yet be established from two 2026 papers alone.
Reproducibility55/100
The reproducibility score is 55, the mean over 2 reviewed papers, each rated medium reproducibility readiness. Evolve's review of the low-bit distillation paper flags single-run evidence and partial ablations despite well-controlled comparisons, while the MoE pruning study is judged convincing, solid empirical work — a mid-range rating overall.
Research summary
Chuangyi Li is a researcher at the Chinese Academy of Sciences working on machine learning for efficient LLM inference and model compression. The Evolve catalog contains two 2026 arXiv papers, on both of which Li is a middle author without corresponding status. The first, 'Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning', shows that on-policy distillation applied after quantization-aware distillation restores sub-3-bit reasoning, lifting MATH retention from 35% to 70%; Evolve calls it well-controlled but notes single-run evidence and partial ablations. The second, 'You Only Need 2/3 of the Chosen Experts', is an empirical study of dynamic expert pruning in fine-grained MoE LLMs, finding that uniform 2/3 expert retention keeps 98.8% of performance while dynamic pruning pays only under aggressive budgets — rated convincing, solid work. Both papers carry medium reproducibility readiness, giving a reproducibility mean of 55. The bibliometric record lists 19 distinct coauthors (collaboration score 60), with Yuanteng Chen, Peisong Wang and Yuantian Shao recurring most often. The field distribution splits evenly between two near-identical efficient-inference labels, and the computed focus score is 0, which should be read cautiously. Output, impact and activity axes remain unscored for lack of data, so this is an early-stage profile built from a small but coherent record in inference efficiency.
Collaboration pattern
Chuangyi Li's record shows 19 distinct coauthors (collaboration score 60). The most frequent collaborators are Yuanteng Chen, Peisong Wang and Yuantian Shao, each appearing twice in the frequent-coauthor list, alongside Jing Liu, Jian Cheng, Qiwei Lai, Chen Tianqi, Nanxin Zeng and Zhilei Liu. On both cataloged 2026 papers Li holds a middle author position without corresponding authorship, indicating contribution within a Chinese Academy of Sciences team rather than sole leadership.
Research keywords
Highlights
On-policy distillation restores sub-3-bit reasoning
Two-thirds of chosen experts keep 98.8% of MoE performance
A consistent empirical line in inference efficiency
Across both cataloged 2026 papers — spanning quantization-aware distillation and MoE expert pruning — the common thread is careful, budget-controlled empirical analysis of LLM inference cost, so far in middle-author roles within a Chinese Academy of Sciences collaboration network.
Timeline
- 2026Enters the Evolve catalog with an empirical study of dynamic expert pruning in fine-grained MoE LLMs, reporting that uniformly keeping 2/3 of the chosen experts retains 98.8% of performance.
- 2026Publishes on-policy distillation for low-bit reasoning, restoring sub-3-bit MATH retention from 35% to 70% after quantization-aware distillation.
- 2026Both arXiv papers receive medium reproducibility-readiness ratings, yielding a reproducibility mean of 55 across 2 reviewed papers, alongside a collaboration score of 60 from 19 distinct coauthors.
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.
- Dimension scores are computed by Evolve's bibliometric radar; the output, impact and activity axes lack sufficient data and are not scored.
- This profile rests on only 2 Evolve-reviewed papers (both 2026 arXiv preprints), so it may not reflect the full publication record.
- No ORCID is provided; the profile is name-resolved from public bibliographic data, and 'Chuangyi Li' may correspond to more than one researcher.
- The two field labels in the raw data are near-duplicates split 0.5/0.5, so the focus score of 0 should be interpreted with caution.