First-author researcher in efficient LLM inference: extreme low-bit quantized reasoning and MoE expert pruning
2
papers here
2
reviewed by Evolve
2
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 yet: the snapshot provides no total publication count, venue list, or per-year output volume for this researcher. This axis would measure cumulative and annual paper output across indexed sources.
Impact—
Impact cannot be scored yet: no citation counts, h-index, or downstream-adoption signals are present in the data. This axis would measure how widely the two quantization and MoE-pruning studies are cited and used.
Focus0/100
The platform reports a focus score of 0/100, which sits oddly against the raw field distribution: 100% of labeled activity falls in one narrow area, split 0.5/0.5 between machine learning (efficient LLM inference) and its model-compression variant. Both cataloged 2026 papers confirm this single-agenda concentration, so the zero score likely reflects label fragmentation rather than demonstrated topic breadth; read the raw distribution as the safer signal.
Collaboration60/100
The collaboration score is 60/100, built on a raw count of 19 distinct coauthors in the bibliometric snapshot. The frequent-coauthor list shows repeated names — Peisong Wang and Yuantian Shao each appear twice — alongside Jian Cheng, Chuangyi Li, Jing Liu, Qiwei Lai, Nanxin Zeng, and Zhilei Liu. Both cataloged papers are first-authored but multi-author, consistent with small-team work rather than solo output.
Activity—
Activity data is insufficient to score: the snapshot covers only two 2026 arXiv entries with no multi-year publication rhythm. This axis would measure ongoing output cadence and recency across the researcher's full record.
Reproducibility55/100
Reproducibility scores 55/100, a mean taken over the two cataloged papers, both tagged medium reproducibility readiness. Evolve's review notes are consistent with that mid-level score: the quantization-distillation paper rests on single-run evidence with partial ablations, while the MoE pruning study is described as convincing, solid empirical work. The signal suggests reproducible but not yet fully open or multi-seed research.
Research summary
Yuanteng Chen is an early-documentation researcher working at the intersection of efficient large language model (LLM) inference and model compression, profiled here from public bibliographic data and two Evolve-reviewed papers. Both cataloged works are 2026 arXiv preprints with Chen in the first-author position (not corresponding author). The first applies on-policy distillation after quantization-aware distillation (QAD), showing that training in the quantized model's own output space restores sub-3-bit reasoning — MATH retention recovers from 35% to 70% — though Evolve notes single-run evidence and partial ablations. The second is an empirical study of dynamic expert pruning in fine-grained mixture-of-experts (MoE) LLMs, finding that uniformly keeping 2/3 of the chosen experts retains 98.8% of performance, with dynamic pruning paying off only under aggressive budgets; Evolve judges it convincing, solid work. The field distribution is tightly concentrated: two adjacent labels covering efficient LLM inference and model compression at 0.5 each. Collaboration signals are moderate (score 60/100) with 19 distinct coauthors and repeat collaborators such as Peisong Wang and Yuantian Shao. Reproducibility averages 55/100 across two papers tagged medium readiness. Output, impact, and activity axes remain unscored for lack of data; no institution, country, or ORCID is attached to this profile.
Collaboration pattern
The snapshot lists 19 distinct coauthors. Peisong Wang and Yuantian Shao each appear twice in the frequent-coauthor list, pointing to stable repeat collaborations, alongside Jian Cheng, Chuangyi Li, Jing Liu, Qiwei Lai, Nanxin Zeng, and Zhilei Liu. Both cataloged papers are first-authored (non-corresponding) multi-author efforts, consistent with leading the technical work inside a small group; the collaboration score is 60/100.
Research keywords
Highlights
Restoring reasoning after extreme quantization
Two-thirds of the chosen experts may be enough
A focused first-authored efficiency agenda
Both 2026 first-authored papers attack complementary deployment costs of LLMs — weight precision via sub-3-bit quantization and expert routing via pruning — with medium reproducibility readiness and a reproducibility mean of 55/100 across the two papers.
Timeline
- 2026Posts a first-authored arXiv paper on on-policy distillation after quantization-aware distillation, restoring sub-3-bit reasoning with MATH retention rising from 35% to 70%.
- 2026Posts a first-authored empirical study showing that uniform 2/3 expert retention keeps 98.8% of fine-grained MoE performance, with dynamic pruning paying off only under aggressive budgets.
- 2026Both papers enter the Evolve catalog with medium reproducibility readiness and a reproducibility mean of 55/100 across the two papers.
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 signals come from a narrow public snapshot (19 distinct coauthors, two Evolve-reviewed papers); no citation counts, h-index, or venue history are available.
- No institution, country, or ORCID is attached; this profile is name-resolved from public bibliographic data, and the name Yuanteng Chen may be ambiguous.
- The reviewed record covers only 2026 arXiv preprints; earlier career history, education, and other publications fall outside the data scope.
- The reported focus score of 0/100 conflicts with a field distribution concentrated entirely in one narrow area (two adjacent labels at 0.5/0.5); treat that axis with caution.