Liangli Zhen

Research profile

Research summary

Liangli Zhen is a researcher at the Agency for Science, Technology and Research (A*STAR), Singapore, whose OpenAlex-resolved record (h-index 25, i10-index 39) splits almost evenly between two demanding areas: machine learning for efficient LLM inference and AI in ophthalmology. Topic modelling of the record places Advanced Neural Network Applications first (27 works), followed by domain adaptation and few-shot learning (19), image and video retrieval (16), multimodal machine learning (12), and adversarial robustness (10). Recent activity is strong: 38 works in the last three years, with annual counts rising from 11 (2024) to a peak of 16 (2025) and 11 already in 2026, supported by 42 distinct coauthors. Two 2026 arXiv preprints reviewed by Evolve illustrate the dual track. In ophthalmology, a multicentre uncertainty-aware vision transformer screened glaucoma across multi-ethnic myopic and non-myopic populations, achieving AUROC 86–100% over 16 datasets — though Evolve notes the headline metrics are aided by excluding uncertain cases. In LLM efficiency, JaCAP extends the authors' linear CAPKV with a softmax-Jacobian information-capacity objective for KV cache eviction; the theory is clean, but the deployed score drops competition terms and gains are untested statistically. Evolve rates reproducibility modestly (mean 38 across 2 papers), and the even 0.5/0.5 field split yields a low focus score — breadth over concentration.

Research keywords

Highlights

• Multicentre glaucoma screening with an uncertainty-aware Vision Transformer

  Across 16 datasets spanning multi-ethnic myopic and non-myopic populations, the model achieved AUROC 86–100% with strong external validation. Evolve cautions that headline metrics are aided by excluding uncertain cases, and rates reproducibility as low.

• JaCAP: Jacobian-based capacity preservation for KV cache eviction

  Extends the authors' linear CAPKV with a softmax-Jacobian information-capacity objective for long-context inference. Evolve finds the theory clean but notes the deployed score drops competition terms and gains are untested statistically; reproducibility is rated medium.

• A rare dual-track portfolio: retinal clinical AI and LLM serving efficiency

  Within the same year (2026), Zhen co-authored both the multicentre glaucoma vision-transformer study and the JaCAP KV-cache work — matching the even 0.5/0.5 field split that defines this profile's breadth and low focus score.

Timeline

1. 2013Earliest works in the provided counts-by-year record: 2 publications, the start of a slowly building record.
2. 2020Output reaches 7 works, up from 5 in 2019, as the record begins to accelerate.
3. 20219 works recorded — the strongest full year in the record to that point.
4. 2024Output climbs to 11 works, opening the most productive stretch of the record.
5. 2025Peak year in the record: 16 works.
6. 202611 works already this year, including two arXiv preprints reviewed by Evolve — the multicentre glaucoma ViT study and JaCAP for KV cache eviction.

Papers in the Evolve catalog

• 2026-09-24

  Detecting Glaucoma Across Multi-ethnic Myopic and Non-Myopic Populations Using an Uncertainty-Aware Vision Transformer: A Multicentre Model Development and Validation Study

• 2026-09-08

  Jacap: Robust KV Cache Eviction via Jacobian-Based Nonlinear Information Capacity Preservation

Frequent co-authors