Alibaba Group researcher working across LLM-agent training, model merging, and audio-visual dialogue systems.
4
papers here
4
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 yet scored: the platform lacks complete publication-count and venue data for this researcher. This axis would measure total scholarly output, venue quality, and the share of first- or corresponding-author papers.
Impact—
Not yet scored: citation counts, h-index, and downstream uptake are unavailable in this snapshot. This axis would measure citations to, and real-world adoption of, the four reviewed 2026 papers.
Focus0/100
The focus score is 0, the floor of the scale: the four reviewed 2026 papers divide evenly — 25% each — across reinforcement learning for language-model agents, parameter-efficient fine-tuning of LLM agents, model merging for multimodal LLMs, and speech-vision multimodal learning. The record shows breadth across the LLM-agent and multimodal stack rather than depth in a single specialization.
Collaboration70/100
Collaboration scores 70/100, supported by 33 distinct coauthors across only four reviewed papers. The recurring roster includes Xiaobing Tu, Rongkun Xue, Xiantao Zhang, Haochen Li, Miancan Liu, Ziyuan Chen, Yinggui Wang, Jinkui Ren, Hongliang Wei, and Debin Zhao. Liu sits in the middle author position on all four papers and is corresponding author on none, consistent with dense, team-based industrial research.
Activity—
Not yet scored: only four 2026 catalog entries exist, so yearly output and recent-activity signals cannot be computed. This axis would measure publication cadence and research continuity over time.
Reproducibility38/100
Reproducibility readiness averages 38/100 across all 4 assessed papers. Two agent-training efforts — OmniVChat and HarnessBandit — reach medium readiness, while TAC-Merge and BQ-LoRA are rated low; Evolve's verdicts attribute part of the gap to missing code releases for the model-merging and fine-tuning methods, so replication currently depends on internal infrastructure.
Research summary
Zhengxi Liu is a researcher affiliated with Alibaba Group (China) whose reviewed work, as indexed by Evolve, concentrates on the training and adaptation of large language models and multimodal systems. The catalog contains four 2026 arXiv papers spread evenly across four subfields — reinforcement learning for language-model agents, parameter-efficient fine-tuning of LLM agents, model merging for multimodal large language models, and multimodal speech-vision learning — each accounting for 25% of the reviewed output. This even spread yields a focus score of 0, reflecting a breadth-first rather than depth-first profile. On every paper Liu is a middle-position author and not the corresponding author, suggesting a contributor role within larger teams; the bibliometric record lists 33 distinct coauthors and a collaboration score of 70. Reproducibility is the weakest measured axis, with a mean readiness of 38 across 4 assessed papers: the two agent-training papers (OmniVChat, HarnessBandit) are rated medium, while TAC-Merge and BQ-LoRA are rated low, partly for absent code releases. Output, impact, and activity metrics are not yet available. All figures derive from Evolve's review of arXiv-indexed work and OpenAlex-based bibliometrics.
Collaboration pattern
Liu's reviewed work is team-based: 33 distinct coauthors across only four 2026 papers, with a collaboration score of 70. Recurring partners include Xiaobing Tu, Rongkun Xue, Xiantao Zhang, Haochen Li, Miancan Liu, Ziyuan Chen, Yinggui Wang, Jinkui Ren, Hongliang Wei, and Debin Zhao. Liu holds a middle author position on all four papers and is corresponding author on none, consistent with contributing specialist roles inside large Alibaba research groups rather than solo-led projects.
Research keywords
Highlights
Native audio-visual dialogue: synthetic data engine plus rubric-based RL
Bandit-scheduled harnesses for agentic RL training
Curvature-guided merging for multimodal LLMs
Quotient-geometry LoRA for LLM agents
Timeline
- 2026OmniVChat: synthesizing, benchmarking, and training for native audio-visual dialogue (arXiv preprint, middle author).
- 2026HarnessBandit: joint learnability-transferability scheduling for multi-harness agentic reinforcement learning (arXiv preprint, middle author).
- 2026TAC-Merge: tracing and coordinating cross-layer influence for multimodal model merging (arXiv preprint, middle author).
- 2026BQ-LoRA: behavior quotient learning for low-rank adaptation of LLM agents (arXiv preprint, middle author).
Papers in the Evolve catalog
OmniVChat: Synthesizing, Benchmarking, and Training for Native Audio-Visual Dialogue
HarnessBandit: Joint Learnability-Transferability Scheduling for Multi-Harness Agentic Reinforcement Learning
Tracing and Coordinating Cross-Layer Influence for Multimodal Model Merging
Behavior Quotient Learning for Low-Rank Adaptation of LLM Agents
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.
- Name is common (Liu/Zhengxi); this profile is name-resolved from public bibliographic data and the reviewed arXiv papers — identity beyond the stated Alibaba Group affiliation is not independently verified.
- Bibliometrics (distinct coauthors, field distribution, radar scores) are computed from OpenAlex-derived data only.
- Output, impact, and activity dimensions are unscored because citation, venue, and yearly-output data were unavailable in this snapshot.
- All four reviewed papers are 2026 arXiv preprints; peer-reviewed venue status and the Chinese-character form of the author name are unverified.