PH.D. STUDENT — INST. OF AI & FUTURE NETWORKS, BNU

Qianli
Ma

~$

I work on efficient LLM systems, semantic state transfer, multi-candidate reasoning, and security-oriented machine learning — building AI that is fast to serve and safe to trust.

01 /

Research

From KV-cache level systems optimization to model-level security analysis — one goal: AI that is fast, cheap, and trustworthy.

[01]

Efficient LLM Inference

KV-cache reuse, semantic cache distillation, state transfer, selective recomputation, and bandwidth-aware serving.

[02]

Reasoning Pipelines

Candidate construction, fixed-verifier reranking, coverage-conversion gaps, and answer-mode compatibility.

[03]

Trustworthy ML

Backdoor attacks, frequency-domain robustness, model behavior under perturbation, and security evaluation.

[04]

Applied Computer Vision

Vision pipelines for smart livestock systems, identity-related signals, weight estimation, and agricultural AI.

02 /

Selected Publications

Recent work on cache distillation, multi-candidate reasoning, and frequency-domain backdoors.

ICML2026

Semantic Cache Distillation: Efficient State Transfer via Reuse and Selective Patching

Qianli Ma, Zhiqing Tang, Hanshuai Cui, Zhi Yao, Weijia Jia

Proc. 43rd International Conference on Machine Learning

NeurIPS2026

Correct but Unselectable: The Hidden Interface Tax in Multi-Candidate Reasoning

Qianli Ma

Under review

IEEE SPL2023

Stealthy Frequency-Domain Backdoor Attacks: Fourier Decomposition and Fundamental Frequency Injection

Qianli Ma, Junping Qin, Kai Yan, Lei Wang, Hao Sun

NNICE2024

Perceptually Imperceptible Backdoor Attacks Using High-Frequency Information in Deep Learning Models

Qianli Ma, Junping Qin, Yin Cao, Jiaqi Ren

IET CV2024

FastFaceCLIP: A Lightweight Text-driven High-Quality Face Image Manipulation

Jiaqi Ren, Junping Qin, Qianli Ma, Yin Cao

03 /

Experience

Systems-oriented research with applied ML depth.

LLM_SERVING

Semantic cache distillation for bandwidth-limited deployment

Designed REUSE and PATCH mechanisms for state transfer between shared-architecture, weight-mismatched models.

AI_SECURITY

Backdoor behavior and frequency-domain attack analysis

Studied stealthy backdoor mechanisms, trigger design, and robustness-oriented model analysis.

SMART_AGRICULTURE

Computer vision for livestock systems

Built applied pipelines for cattle image acquisition, weight estimation, visual signal analysis, and farm decision support.

04 /

Open to research collaboration.

mql16899@163.com