A-ZnkQianli Ma

Master's student at USTC · AI safety and large language model systems

Between model behavior and real systems,looking for answers we can verify.

My goal is to turn large language models that answer questions into systems that complete tasks.

I work on large language model systems, model safety, and agent engineering, with an emphasis on turning research into working systems. My research on multimodal large language models and safety spans adversarial attacks, privacy protection, and representation analysis, extending these questions into evaluation harnesses, behavior control, and system prototypes.

INDEX / Research dossier

Research

Research tracks

3 tracks

LLM / MLLM systems

Turning model capabilities into system capabilities through multimodal reasoning, research harnesses, and evaluation loops.

  • Multimodal large language models (MLLMs)
  • Inference frameworks, evaluation systems, and deployment
  • Reproducible experiments and systematic validation
Three drawings of cerebral cortex tissue by Cajal, showing cells and nerve fibers arranged across layers.
Layers and connections of the cortex1899
Wikimedia Commons
About this plate

From Cajal's comparative studies of the human cerebral cortex. Layers, connections, and local structures offer a historical perspective for these systems research notes.

Santiago Ramón y CajalPublic domain; original work from 1899, digital composite assembled in 2009.

Model safety

Studying model vulnerabilities in open environments and how to make safety mechanisms transferable and controllable.

  • Adversarial examples and attacks
  • Safety alignment and refusal behavior control
  • Model fingerprinting and copyright protection
A wireframe cube with diagonals, alongside two spatial interpretations emphasized by heavier lines.
The same lines, different interpretations1899
Popular Science Monthly / Commons
About this plate

Compare the heavy lines in the middle and right drawings, then return to the left. The same visual cues support different spatial interpretations—a historical reference for uncertainty in perception and judgment.

Illustration in Joseph Jastrow's The Mind's Eye; artist uncreditedPublic domain; figures 13, 13a, and 13b in the 1899 publication.

Agent engineering

Moving beyond one-off prompts toward maintainable, measurable, and iterative systems engineering.

  • Prompt Engineering → Harness Engineering
  • Multi-agent collaboration systems
  • Methods for rapidly building MVPs
Mechanical linkage between a centrifugal governor and a throttle valve, retaining the original component labels and English caption.
Bring feedback into the system1900
Routledge / Wikimedia Commons
About this plate

The governor feeds changes in rotational speed back into valve control. Sensing, feedback, and action form a closed loop, providing a tangible reference for feedback in systems engineering.

Historical illustration from a work by R. RoutledgePublic domain; illustration from the 13th edition of Discoveries & Inventions of the Nineteenth Century.

Selected Publications

Papers and research

Published and accepted papers, with their research questions, methods, and experiments.

02 papers

IEEE TIFSPublished

DUAP: Disentanglement-Based Universal Adversarial Perturbations for Robust Multilingual Speech Privacy Protection

Universal multilingual adversarial perturbations for speech privacy against Whisper

Language feature disentanglement and gradient optimization produce transferable perturbations that reduce Whisper’s ability to transcribe sensitive speech.

DUAP Figure 3: language and content encoders, feature-swapping reconstruction, Whisper, and three loss functions.
Fig. 3

Language feature disentanglement (LFD): separating language and content features, then reconstructing exchanged features to generate Mel spectrograms with target-language features.

↗
View all papers and research notes

Practice

Projects / Practice

Research harnesses, behavior steering, content systems, and workflows.

4 items
EXPERIMENT / Research harness

Multimodal model research harness

An experiment harness built around Qwen2.5-VL for batch evaluation, safety experiments, and reproducible research.

Related content
CONTROL / Behavior steering

Behavior control with steering vectors

Steering-vector methods for adjusting model behavior and output style, investigating controllable generation and refusal boundaries.

Related content

Lens

Current questions and writing scope

OPEN QUESTIONS / Research

Current questions

  • How to build agent systems that can be verified and evaluated
  • How to build transferable safety mechanisms for multimodal models
  • How engineering harnesses can address fragile prompts
FIELDNOTES / Writing scope

What I write about

  • Large language model systems and safety
  • Agent engineering and workflow design
  • AI coding and prompt engineering methods
  • Project retrospectives and technical reflections

Social

Elsewhere

Find me on these platforms. Public statistics refresh where supported.

4 platforms

Contact

Contact / Links

Academic profiles, open-source code, and email.

Writing

Writing

Published articles on this blog by A-Znk.

4 posts

Image preview

Open original image