Alibaba researcher spanning LLM-agent RL, parameter-efficient fine-tuning, agent memory, and multimodal model merging
4
papers here
4
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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—
Output data is insufficient: the OpenAlex record lists h-index 0 and no year-by-year publication counts, so productivity cannot be scored. This axis would measure total works, yearly output volume, and authorship position across the record.
Impact—
Impact cannot be assessed: citation counts and citing works are absent from the available snapshot (i10-index reads 0), so influence is unscored. This axis would measure citations, h-index trajectory, and uptake of the arXiv preprints.
Focus0/100
The focus score is 0, consistent with the field distribution: four directions — RL for language-model agents, parameter-efficient fine-tuning of LLM agents, LLM-agent memory systems, and model merging for multimodal LLMs — each hold exactly 0.25 weight, an even spread with no dominant research line yet.
Collaboration64/100
Score 64 rests on 24 distinct coauthors in the recent record plus a recurring core cluster — Xiaobing Tu, Zhengxi Liu, Rongkun Xue, Jinkui Ren, Haochen Li, and Yinggui Wang appear repeatedly in the coauthor list — indicating steady team-based collaboration at Alibaba rather than a broad independent network.
Activity61/100
Score 61 reflects 7 works over the last three years, with the counts-by-year record placing all 7 in 2026 — a concentrated burst containing the four Evolve-cataloged arXiv preprints across agent RL, LoRA fine-tuning, memory management, and model merging. Publishing is active, but the visible window is short.
Reproducibility38/100
Score 38 is the reproMean across 4 reviewed papers: Evolve rates HarnessBandit and MemForest at medium readiness, and TAC-Merge and BQ-LoRA at low. Verdicts repeatedly cite missing code releases, absent signal ablations, no significance tests, and confounded attribution — mathematical framing runs ahead of empirical verification.
Research summary
Xiantao Zhang is a researcher at Alibaba Group (China) whose recent work concentrates on the systems side of large language model agents. The 2026 topic distribution shows an even four-way split: reinforcement learning for language-model agents, parameter-efficient fine-tuning of LLM agents, LLM-agent memory systems, and model merging for multimodal LLMs, each holding 0.25 weight. Four arXiv preprints, all dated 2026, anchor the Evolve catalog: HarnessBandit schedules training harnesses per GRPO step with a multi-armed bandit; TAC-Merge traces cross-layer multimodal influence with curvature graphs to guide model merging; BQ-LoRA compresses redundant agent-trajectory updates under a fixed LoRA rank using quotient geometry; and MemForest partitions agent memory into EventTrees for progressive merging, claiming 50% compression at 97-99.7% retention. Evolve's verdicts credit the mathematical framing across all four while flagging thin empirical validation — single small runs, missing signal ablations, no significance tests, and no code releases for two papers. With 7 works in the past three years and 24 distinct coauthors, this is an active, early-visibility industrial profile. The record is name-resolved from OpenAlex public bibliographic data; citation-based output and impact measures are not yet available.
Collaboration pattern
The record shows 24 distinct coauthors with a dense recurring core: Xiaobing Tu, Zhengxi Liu, Rongkun Xue, Jinkui Ren, Haochen Li, Miancan Liu, Ziyuan Chen, Yinggui Wang, Hongliang Wei, and Debin Zhao — Jinkui Ren and Yinggui Wang appear twice in the frequent-coauthor list. Across the four cataloged papers Zhang holds middle or last author position and is never corresponding author, consistent with specialist contributions inside an Alibaba industry team.
Research keywords
Highlights
HarnessBandit: bandit-scheduled multi-harness agentic RL
TAC-Merge: curvature-guided multimodal model merging
BQ-LoRA: quotient geometry for LLM-agent fine-tuning
MemForest: EventTree memory compression for agents
Timeline
- 2026Publishes HarnessBandit, scheduling training harnesses per GRPO step with a multi-armed bandit for multi-harness agentic reinforcement learning.
- 2026Publishes TAC-Merge on curvature-graph tracing of cross-layer influence for multimodal model merging, and BQ-LoRA on quotient-geometry LoRA adaptation of LLM agents.
- 2026Publishes MemForest, proposing EventTree partitioning and progressive merging for agent memory with claimed 50% compression at 97-99.7% retention.
- 2026All four preprints enter the Evolve catalog, with reproducibility rated medium for two and low for two; recent3yWorks reaches 7 across 24 distinct coauthors.
Papers in the Evolve catalog
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
MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging
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.
- Bibliometrics come from an OpenAlex-derived snapshot only; output and impact axes are missing because citation counts are absent (h-index and i10-index read 0).
- No ORCID is attached; the profile is name-resolved from public bibliographic data and may merge or split namesakes of a relatively common name.
- Coverage is limited to 4 arXiv preprints reviewed by Evolve; peer-reviewed publications, if any, fall outside this snapshot.
- Reproducibility scores and verdicts reflect Evolve's review of these preprints only, not independent replication.