Yuan Liu — Alibaba researcher on RL-driven LLM efficiency (speculative decoding) and agent post-training alignment
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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—
Not scored — no publication-volume data is available in this snapshot. This axis would measure total papers per year and venue coverage; only two reviewed 2026 arXiv papers exist here, too thin to estimate output.
Impact—
Not scored — no citation, usage or downstream-adoption metrics are provided. This axis would measure influence such as citations and independent reproductions; both papers appeared in 2026, so no such evidence exists yet.
Focus0/100
Scores 0/100 on this axis. The bibliographic field distribution splits exactly 50/50 between machine learning / efficient LLM training systems and machine learning / NLP (LLM agent post-training alignment), and the two reviewed 2026 papers sit one in each — GrowMTP on speculative-decoding draft heads inside the RL loop, tlm-DRE on turn-level credit assignment for agents — indicating two parallel directions rather than one concentrated research agenda.
Collaboration44/100
Scores 44/100 with 8 distinct coauthors recorded across the reviewed 2026 work. Recurring partners include Minghua He, Lingzhe Zhang, Xiao Zhou, Aiwei Liu, Zishuo Zhao, Kai Chen and Ao Li, suggesting a stable team. On both papers Yuan Liu sits in middle or last author position and is never the corresponding author, indicating contribution within a group rather than an independently led research line.
Activity—
Not scored — no yearly publication volume or multi-year timeline is available beyond the two 2026 papers. This axis would measure research cadence and continuity across consecutive years, which this snapshot cannot capture.
Reproducibility38/100
Scores 38/100, the mean across the two papers carrying reproducibility assessments (reproPapers: 2). GrowMTP rates medium: results are convincing at 4–7B scale but come from a single run, with a modest 9.4% compute saving. tlm-DRE rates low: gains over DIL are small, weakly tested and internally inconsistent. Neither paper clears a high reproducibility bar, so this early mean sits well below the midpoint.
Research summary
Yuan Liu is a researcher at Alibaba Group (China) whose reviewed work spans two branches of modern large language model research: efficient LLM training systems and post-training alignment for LLM agents. The Evolve catalog contains two 2026 arXiv papers. The first, GrowMTP, explores whether reinforcement learning can grow its own speculative-decoding draft head: a draft head is trained from scratch inside the RL loop using verification signals, with convincing results at 4–7B scale, though from a single run and with a modest 9.4% compute saving. The second, on turn-level multiscale density ratio estimation (tlm-DRE), adds turn-confidence weighting to DRE-based imitation learning for LLM agents; Evolve's assessment finds the gains over DIL small, weakly tested and internally inconsistent. Bibliometrically, the picture is early-stage: 8 distinct coauthors are recorded, with recurring partners including Minghua He, Lingzhe Zhang, Xiao Zhou, Aiwei Liu, Zishuo Zhao, Kai Chen and Ao Li; authorship sits in middle or last position, without a corresponding-author role in either paper. The field distribution splits evenly (50/50) between the two directions, producing a focus score of 0/100, while the reproducibility mean across the two assessed papers is 38/100. Output, impact and activity axes currently lack underlying data and are not scored. 'Yuan Liu' is a common name, so this profile is name-resolved from public bibliographic data.
Collaboration pattern
The record shows a moderate, stable collaboration footprint: 8 distinct coauthors across the two 2026 papers, giving a collaboration score of 44/100. Recurring names include Minghua He, Lingzhe Zhang, Xiao Zhou, Aiwei Liu, Zishuo Zhao, Kai Chen and Ao Li. Yuan Liu sits in middle or last author position on both papers and is not corresponding author on either, consistent with a senior contributor within a team rather than a solo-led line.
Research keywords
Highlights
GrowMTP: RL grows its own speculative-decoding draft head
tlm-DRE: turn-level density ratio estimation for LLM agents
Timeline
- 2026Last-author arXiv paper 'Turn-level Multiscale Density Ratio Estimation for LLM Agents' appears, adding turn-confidence weighting to DRE imitation learning for LLM agents.
- 2026Middle-author arXiv paper 'GrowMTP: Can RL Grow Its Own Draft Head?' appears, training a speculative-decoding draft head from scratch inside the RL loop.
- 2026Both papers enter the Evolve catalog in the same year, splitting evenly across efficient LLM training systems and LLM agent post-training alignment, with a reproducibility mean of 38/100.
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.
- 'Yuan Liu' is a common name; this profile is name-resolved from public bibliographic data, and the frequent-coauthor list even contains a same-name entry, which may be a namesake or a data artifact.
- Radar scores rest on a very small base: only 2 papers carry reproducibility assessments (reproMean 38, reproPapers 2), so scores are early-stage and volatile.
- Output, impact and activity axes lack underlying bibliometric data in this snapshot and are left unscored rather than estimated.
- Both catalog papers are 2026 arXiv preprints; no peer-reviewed venue, ORCID or country information was available.