Curriculum Vitae — Kainan Liu

Education

The Hong Kong University of Science and Technology (Guangzhou) 2023 - 2025

M.Sc. in Data-Centric Artificial Intelligence

Zhongnan University of Economics and Law 2019 - 2023

B.Sc. in Data Science and Big Data Technology

Publications

Astra: Activation-Space Tail-Eigenvector Low-Rank Adaptation of Large Language Models

ACL 2026 Findings

Kainan Liu^*, Yong Zhang^*, Ning Cheng^†, Yun Zhu, Yanmeng Wang, Shaojun Wang, Jing Xiao

Proposes a novel PEFT method that estimates the activation covariance matrix from a task-specific calibration set (as few as 64 samples), performs eigendecomposition to extract the tail eigenvector subspace, and constrains low-rank updates within this weakly-activated subspace to avoid interfering with the pretrained model's dominant representations. Evaluated on 16 benchmarks spanning GLUE, mathematical reasoning, code generation, and commonsense reasoning, Astra significantly outperforms LoRA, DoRA, PiSSA, LoRA-GA and other PEFT methods — achieving substantial gains on CoLA and MRPC where standard LoRA collapses — and surpasses full fine-tuning in certain scenarios. Code open-sourced.

GRASP: Replace Redundant Layers with Adaptive Singular Parameters for Efficient Model Compression

EMNLP 2025 Main Conference

Kainan Liu^*, Yong Zhang^*, Ning Cheng^†, Zhitao Li, Shaojun Wang, Jing Xiao

A gradient-guided hybrid compression framework that combines redundant layer identification with adaptive singular parameter selection. It locates functionally redundant Transformer layers via cosine similarity of hidden states between adjacent layers, then uses gradient-attributed importance scores (rather than magnitude-based heuristics) to adaptively retain critical singular components from SVD, replacing full layers with lightweight parameterizations. At 20% compression, it preserves ~90% of original performance across 19 datasets spanning 5 model families (LLaMA, LLaMA 2/3, Mistral-7B), outperforming SliceGPT, LaCo, LLM-Pruner, and others. The compression process is training-free and completes in ~0.16 hours on a single A100 GPU.

Detecting and dissecting anomalous anatomic regions in spatial transcriptomics with STANDS

Nature Communications

Kaichen Xu^†, Yan Lu, Suyang Hou, Kainan Liu, Yihang Du, Mengqian Huang, Hao Feng, Hao Wu, Xiaobo Sun^*

A GAN-based multi-task deep learning framework that jointly models gene expression and histology image information for detection, cross-sample alignment, and subtyping of anomalous tissue domains in spatial transcriptomics data. It integrates multi-modal features via graph attention networks and Transformer fusion modules, validated across multiple platforms including 10x Visium, Slide-seqV2, and Stereo-seq. STANDS can detect early-stage cancerous regions visually indistinguishable in histology images, and is the only method that accurately performs cross-sample alignment and anomaly subtyping in multi-sample settings.

^* Equal contribution ^† Corresponding author

Experience

Ping An Technology (Shenzhen) Co., Ltd. Aug 2025 - Present

AI Researcher & Algorithm Engineer

Training and optimizing LLMs for financial risk control and corporate banking analysis, including domain-adaptive fine-tuning, instruction tuning, and prompt optimization to improve financial NLP task performance
Proposed Astra (ACL 2026 Findings): a low-rank adaptation method based on activation covariance tail-eigen subspaces, requiring only 64 calibration samples to build task-adaptive adapters, outperforming existing PEFT methods on 16 NLU/NLG benchmarks
Building evaluation and monitoring pipelines to track model performance across financial NLP benchmarks

Ping An Technology (Shenzhen) Co., Ltd. Sept 2024 – Aug 2025

Research Intern

Conducted research on model compression techniques, focusing on identifying and replacing redundant Transformer layers with parameter-efficient alternatives
Proposed GRASP (EMNLP 2025 Main Conference): a hybrid compression framework combining cosine-similarity-based redundant layer identification with gradient-guided singular value selection, achieving ~90% of original performance at 20% compression across models from LLaMA-7B to LLaMA 3.1-8B
Implemented and systematically evaluated various pruning and compression methods on LLaMA 2-7B/13B, LLaMA 3.1-8B-Instruct, and Mistral-7B, covering 19 benchmark datasets

Vipshop (China) June 2024 – Sept 2024

Recommendation Algorithm Intern

Processed and analyzed large-scale user-item interaction data and behavioral sequences for recommendation system enhancement
Deployed XGBoost models for recommendation ranking optimization, improving key engagement metrics
Developed feature engineering pipelines for extracting user behavior patterns from historical interaction data

Technical Skills

Programming Python Frameworks PyTorch, PEFT, Transformers, vLLM Research Areas PEFT, Model Compression, AI Agents Languages Chinese (native), English (IELTS 7.0)