CAS researcher making LLMs cheaper to run: low-bit quantization, on-policy 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 output score yet: the snapshot lacks this researcher's publication counts. The axis would measure total papers, yearly productivity, and venue spread; the two catalog papers are currently the only evidence in scope.
Impact—
Impact is unscored because citation counts, h-index, and citing works are absent from the snapshot. It would measure uptake of the two 2026 arXiv preprints — both too recent for stable citation signals.
Focus0/100
Focus is recorded at 0/100 in this snapshot. The raw field distribution splits evenly — 0.5 and 0.5 — between 'machine learning (efficient llm inference)' and 'machine learning — efficient llm inference and model compression', two labels describing essentially the same narrow area. The axis measures topical concentration, so the zero appears to reflect label fragmentation in the data rather than genuinely scattered output; interpret this score with caution.
Collaboration60/100
Collaboration scores 60/100 with 19 distinct coauthors in the snapshot, consistent with team-based research at the Chinese Academy of Sciences. Both catalog papers carry middle-position, non-corresponding authorship, fitting a contributor role inside a larger group. The provided frequent-coauthor list highlights repeated partnerships — Peisong Wang and Yuanteng Chen each appear twice — though it also contains duplicates and the researcher's own name, a data-quality artifact worth noting.
Activity—
No activity score: the snapshot provides no per-year publication volume beyond the two 2026 entries. The axis would measure recent output cadence and whether research tempo is sustained over time.
Reproducibility55/100
Reproducibility averages 55/100 across the two papers with readiness evidence (reproPapers 2), and both are individually rated medium. Evolve's verdict on the distillation paper notes well-controlled experiments but single-run evidence and partial ablations; the MoE pruning study is called convincing and solid, built on an iso-budget evaluation design. This suggests careful empirical practice not yet backed by multi-run or open-artifact verification.
Research summary
Yuantian Shao is a researcher affiliated with the Chinese Academy of Sciences whose catalogued work centers on making large language models cheaper to run. The Evolve catalog holds two 2026 arXiv preprints, both with middle-position, non-corresponding authorship. The first trains the student where the quantized model actually operates: after quantization-aware distillation, on-policy training restores sub-3-bit reasoning and lifts MATH retention from 35% to 70% — evidence Evolve judges well-controlled but single-run, with partial ablations. The second is an empirical study of dynamic expert pruning in fine-grained MoE LLMs, finding that uniformly keeping 2/3 of the chosen experts preserves 98.8% of MoE performance while dynamic pruning pays off only under aggressive budgets; Evolve calls it convincing, solid work. Together the papers sketch a coherent efficiency agenda spanning extreme low-bit quantization, distillation, and expert-level sparsity. The radar scores collaboration at 60, with 19 distinct coauthors, and reproducibility at 55 across two medium-readiness papers; output, impact, and activity lack sufficient data, and focus is recorded at 0 even though both field labels describe the same narrow topic. Recurring collaborators in the provided list include Peisong Wang, Yuanteng Chen, and Jian Cheng.
Collaboration pattern
The snapshot records 19 distinct coauthors (collaboration score 60), pointing to steady group-based work at the Chinese Academy of Sciences rather than solo output. Peisong Wang and Yuanteng Chen each appear twice in the provided frequent-coauthor list, alongside Chuangyi Li, Jing Liu, Jian Cheng, Qiwei Lai, Chen Tianqi, Nanxin Zeng, and Zhilei Liu — names reported as given, since the list also contains duplicates and the researcher's own name.
Research keywords
Highlights
On-policy distillation restores reasoning at sub-3-bit quantization
Keeping 2/3 of chosen experts preserves 98.8% of MoE performance
A coherent efficiency agenda across quantization and sparsity
Both 2026 catalog papers attack inference cost from complementary angles — extreme low-bit quantization with distillation, and expert-level sparsity in fine-grained MoE — each with controlled or iso-budget evaluations; both are rated medium repro-readiness.
Timeline
- 2026Catalogued arXiv preprint on on-policy distillation for low-bit reasoning, with middle-position authorship (non-corresponding).
- 2026arXiv empirical study of dynamic expert pruning in fine-grained MoE LLMs, also middle-position authorship.
- 2026Catalog footprint in 2026: two efficiency-focused preprints, 19 distinct coauthors on record, and a reproducibility average of 55.
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.
- Statistics are limited to the provided bibliometric snapshot and the two Evolve-reviewed arXiv papers; no citation counts, h-index, or venue records are in scope.
- Output, impact, and activity axes are unscored because the snapshot lacks publication volumes and citation data for this researcher.
- Both catalog entries are 2026 arXiv preprints with middle-position, non-corresponding authorship; peer-reviewed venues and first/corresponding-author work are not represented in this data.
- The frequent-coauthor list contains duplicate entries and the researcher's own name; the profile is name-resolved from public bibliographic data tied to ORCID 0000-0003-1116-6252.