Alibaba Group researcher on efficient LLM agents, model merging, and accelerated generative inference
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 volume and venue quality cannot be scored yet: the snapshot records an h-index and i10-index of 0 and per-year counts only for 2026 (11 works), so this axis — total publications and the venues they appear in — lacks sufficient data.
Impact—
Citation-based impact is not yet measurable: h-index and i10-index are logged as 0 and no citation counts are attached to the recorded works, so this axis — citations, h-index, and influence on follow-up research — remains unscored.
Focus0/100
The focus score is 0 because the portfolio is evenly dispersed: fieldDist splits 0.25/0.25/0.25/0.25 across efficient generative-model inference, parameter-efficient LLM-agent fine-tuning, LLM-agent memory systems, and multimodal model merging, with no dominant direction. OpenAlex topics echo the spread — Advanced Neural Network Applications (7), GANs and image synthesis (5), domain adaptation and few-shot learning (4), multimodal applications (4), topic modeling (2). This reads as breadth across adjacent efficiency problems rather than narrow specialization.
Collaboration68/100
Scored 68 from 30 distinct coauthors across 11 recent-three-year works, with a stable recurring cluster of 12 names (including Xiaobing Tu, Xiantao Zhang, Linfeng Zhang, and Chang Zou). All four Evolve-reviewed 2026 papers place him in a middle authorship position with no corresponding-author role, consistent with a contributing member of a larger industrial lab team rather than a solo lead.
Activity72/100
Scored 72: the OpenAlex counts-by-year data records 11 works, all in 2026, matching the 11 recent-three-year works. Every Evolve-reviewed paper is a 2026 arXiv preprint spanning four parallel directions — diffusion-transformer caching, agent memory, LoRA agent tuning, and multimodal merging — indicating a burst of near-simultaneous output. Because the record concentrates in a single year, a longer-run cadence cannot yet be verified.
Reproducibility38/100
Scored 38, the mean across 4 Evolve-reviewed papers. Two are rated low readiness — TAC-Merge and BQ-LoRA ship no code — while MemForest and the Gaussian-process feature-cache paper reach medium readiness, with Evolve flagging missing significance tests and abstract-versus-table mismatches. No independent replication evidence exists in scope, so releasing artifacts is the main lever for improving this axis.
Research summary
Jinkui Ren is a researcher at Alibaba Group (China) whose recorded work concentrates on making large-model systems cheaper to run and to adapt. In the Evolve catalog he appears as a middle-position co-author on four 2026 arXiv preprints spanning four equally weighted directions (0.25 each in the field distribution): training-free acceleration of diffusion transformers via a Gaussian-process rectified feature cache, EventTree-based agent memory management (MemForest), quotient-geometry LoRA fine-tuning for LLM agents (BQ-LoRA), and curvature-guided multimodal model merging (TAC-Merge). The OpenAlex snapshot — the profile is name-resolved from that source, with no ORCID link — records 11 works in the recent three-year window, all logged in 2026, alongside 30 distinct coauthors and a recurring cluster of 12 frequent collaborators. Topic assignments (Advanced Neural Network Applications 7, GANs and image synthesis 5, domain adaptation and few-shot learning 4, multimodal applications 4, topic modeling 2) suggest a broader machine-learning base behind the current agent-and-efficiency focus. Evolve's reproducibility assessment averages 38 over the four papers, with two rated low readiness for missing code releases. With h-index and i10-index recorded as 0 and per-year counts limited to a single year, output and impact are deliberately left unscored; the profile is best read as an early-stage view of an active industrial researcher.
Collaboration pattern
All four catalog papers position Jinkui Ren as a middle author with no corresponding role, within a stable team of 12 recurring coauthors (including Xiaobing Tu, Xiantao Zhang, Linfeng Zhang, and Chang Zou) drawn from 30 distinct collaborators over 11 recent works — a collaboration score of 68 that reads as a contributing member of a larger Alibaba lab group rather than a solo line lead.
Research keywords
Highlights
GPR-rectified feature cache for faster diffusion transformers
MemForest compresses agent memory via EventTree partitioning
Quotient-geometry LoRA for LLM agents
Curvature-guided merging for multimodal LLMs
Timeline
- 2026arXiv preprint on accelerating diffusion transformers with a Gaussian-process rectified feature cache (uncertainty-gated recomputation, training-free).
- 2026MemForest preprint proposes EventTree partitioning and progressive merging for agent memory, reporting 50% compression at 97–99.7% retention.
- 2026BQ-LoRA preprint applies quotient geometry to compress redundant agent-trajectory updates under a fixed LoRA rank.
- 2026TAC-Merge preprint traces cross-layer multimodal influence with curvature graphs to tune merge weights via trust-region control.
- 2026OpenAlex logs 11 works in the recent three-year window, all recorded in 2026, with 30 distinct coauthors.
Papers in the Evolve catalog
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
Accelerating Diffusion Transformers with Gaussian Process Rectified Feature Cache
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 snapshot only; no ORCID is linked, so identity is name-resolved from public bibliographic data.
- h-index and i10-index are recorded as 0 and countsByYear covers only 2026 (11 works), so the output and impact axes are unscored.
- Evolve assessments cover four 2026 arXiv preprints with middle authorship and no corresponding-author record; peer-reviewed venues are out of scope.
- The reproducibility score (mean 38 over 4 papers) reflects Evolve's code-release and readiness checks, not independent replication.