Jing Liu

—

Unclaimed · AI-structured profileOpenAlex

Jing Liu — researcher in efficient LLM inference: low-bit quantized reasoning recovery and MoE expert 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—

No publication-volume or venue data underlies this axis yet, so output is not scored. It would measure total output over time — paper counts, venues, and authorship roles. Current evidence covers only two 2026 arXiv preprints.

Impact—

No citation counts, h-index, or influence measures are available, so impact is not scored. This axis would capture citations and downstream adoption. The two reviewed preprints are recent 2026 arXiv postings, too new for meaningful citation evidence.

Focus0/100

The focus score is 0, the tightest value on this axis, and the raw field distribution explains it: weight splits 0.5/0.5 between 'machine learning (efficient llm inference)' and 'machine learning — efficient llm inference and model compression' — two labels describing one domain. Both catalog papers sit inside this same space, covering quantization-aware distillation for low-bit reasoning and expert pruning in fine-grained mixture-of-experts models.

Collaboration60/100

The collaboration score is 60, backed by 19 distinct coauthors in the bibliographic record. Frequent collaborators cluster around a stable core — Yuanteng Chen, Peisong Wang, Yuantian Shao, Chuangyi Li, Jian Cheng, Qiwei Lai, Nanxin Zeng, and Zhilei Liu — several of whom recur across both reviewed papers, indicating an ongoing group rather than scattered one-off ties. On both papers Jing Liu sits in a middle author position without corresponding-author status.

Activity—

No year-by-year publication record exists beyond the two 2026 papers, so activity is not scored. It would measure publishing cadence and continuity across recent years. Both catalog entries are 2026 arXiv preprints; earlier output cannot be confirmed.

Reproducibility55/100

The reproducibility score of 55 is a mean computed over exactly 2 reviewed papers (reproMean 55, reproPapers 2), each rated 'medium' readiness. Evolve's verdicts temper confidence: the distillation work is well controlled but rests on single-run evidence with partial ablations, while the MoE pruning study is judged convincing and solid in its empirical design.

Research summary

Jing Liu is a machine learning researcher working on efficient large language model (LLM) inference, with a catalog footprint concentrated entirely in model compression and sparsity. The Evolve catalog holds two 2026 arXiv preprints with Liu in a middle author position, neither as corresponding author. The first, 'Train Where the Quantized Model Goes', addresses the damage extreme low-bit quantization does to reasoning: after quantization-aware distillation, on-policy distillation lifts MATH retention from 35% to 70%, though Evolve notes single-run evidence and partial ablations. The second, an empirical study of dynamic expert pruning in fine-grained MoE LLMs, finds that uniformly keeping 2/3 of chosen experts retains 98.8% of performance, with dynamic pruning paying off only under aggressive budgets — a verdict Evolve calls convincing and solid. Bibliometric scores place focus at 0, reflecting near-total topical concentration across two field labels that describe the same domain; collaboration at 60 with 19 distinct coauthors; and reproducibility at 55 averaged over the 2 reviewed papers, both rated medium readiness. Frequent coauthors include Yuanteng Chen, Peisong Wang, Yuantian Shao, Jian Cheng, Chuangyi Li, and others. Because 'Jing Liu' is a very common name and no ORCID, institution, or country is attached, this profile is name-resolved from public bibliographic data, and the output, impact, and activity axes remain unscored for lack of data.

Collaboration pattern

Bibliometrics record 19 distinct coauthors and a collaboration score of 60. The frequent-coauthor list clusters around a stable core — Yuanteng Chen, Peisong Wang, Yuantian Shao, Chuangyi Li, Jian Cheng, Qiwei Lai, Nanxin Zeng, Zhilei Liu, and Chen Tianqi — with several names recurring across both reviewed papers, pointing to an ongoing group rather than scattered one-off ties. On both papers Jing Liu appears in a middle author position without corresponding-author status, consistent with a contributing member of a continuing team.

Research keywords

efficient LLM inferencemodel compressionextreme low-bit quantizationquantization-aware distillationon-policy distillationmixture-of-expertsdynamic expert pruningexpert selection redundancyexposure biasLLM reasoningreinforcement learning with verifiersiso-budget evaluation

Highlights

On-policy distillation recovers sub-3-bit reasoning
Evolve's review reports that after quantization-aware distillation (QAD), on-policy distillation restores sub-3-bit reasoning, lifting MATH retention from 35% to 70% — while cautioning that the evidence is single-run and ablations are partial.
Training where the quantized model actually goes
The paper positions on-policy distillation as a second stage that trains on the quantized model's own behavior rather than off-policy teacher trajectories, targeting exposure bias, with reinforcement learning with verifiers among its techniques for LLM reasoning.
Keeping 2/3 of chosen experts retains 98.8% of MoE performance
An empirical study of dynamic expert pruning in fine-grained MoE LLMs, built around iso-budget evaluation, finds uniform 2/3 expert retention keeps 98.8% of performance; dynamic pruning pays only under aggressive budgets. Evolve calls it convincing, solid work.

Timeline

2026Publishes 'Train Where the Quantized Model Goes: On-Policy Distillation for Low-Bit Reasoning' on arXiv; Evolve reports MATH retention recovering from 35% to 70% after on-policy distillation following quantization-aware distillation.
2026Publishes 'You Only Need 2/3 of the Chosen Experts', an empirical study of dynamic expert pruning in fine-grained MoE LLMs, showing uniform 2/3 expert retention keeps 98.8% of performance.
2026Both catalog entries land in this same year, keeping the reviewed record fully inside efficient LLM inference and model compression.
2026Evolve reviews both preprints, rating each at medium reproducibility readiness; the two reviews together form the reproducibility mean of 55.

Papers in the Evolve catalog

Frequent co-authors

Yuanteng Chen×2 Peisong Wang×2 Yuantian Shao×2 Chuangyi Li×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.

Name-resolved from public bibliographic data: 'Jing Liu' is a very common name, and no ORCID, institution, or country is attached, so identity should be verified before reuse.
Bibliometric dimensions are computed from an OpenAlex-based public bibliographic snapshot only; no citation counts, h-index, or venue-tier data underlie the unscored output and impact axes.
The reproducibility score of 55 averages only 2 reviewed papers, both 2026 arXiv preprints, so the evidence base is thin.
Both catalog papers list Jing Liu as a middle author, not corresponding author; claims beyond authorship position cannot be made.