CS598HT: Consistency-Aware Graph Memory for Training-Free VLM Test-Time Adaptation
Urbana, IL
Spring 2026
Member of a three-person team
- Framework: Proposed CAGM, a training-free test-time adaptation framework for vision-language models that combines dynamic memory with graph-based label propagation.
- Consistency Score: Designed a Jensen-Shannon-divergence-based consistency score to measure agreement between zero-shot CLIP predictions and label-propagation predictions.
- Adaptive Mechanism: Used the consistency score as a memory update gate, label-propagation edge weight, and adaptive prediction interpolation coefficient to reduce noisy pseudo-label propagation.
- Evaluation: Implemented the full adaptation pipeline and evaluated the method on standard OOD and cross-domain benchmarks against recent memory- and graph-based TTA baselines.
CS543: Where Do Modern Monocular Depth Models Fail? A Cross-Domain Robustness Study
Urbana, IL
Spring 2026
Member of a three-person team
- Evaluation Pipeline: Built a unified evaluation pipeline to study the robustness of modern monocular depth estimation models under domain shifts such as low light, transparent objects, and scene variation.
- Model Setup: Set up and ran pretrained Depth Anything V2 and ZoeDepth models, including baseline inference and depth-map visualization.
- Benchmarking: Compared Depth Anything V2, ZoeDepth, and Marigold using synthetic RGB-D samples with known ground-truth depth, and computed affine-aligned AbsRel as an initial sanity-check metric.
- Analysis: Analyzed failure patterns across different visual conditions, showing that transparent and specular objects are consistently challenging for all three models.
CS546: Strategic Persuasion in LLMs: Multi-Turn Dialogue with Reinforcement Learning
Urbana, IL
Fall 2025
Member of a six-person team
- Agent Design: Developed a planner-persuader LLM agent for multi-turn strategic persuasion, decoupling high-level persuasion strategy selection from natural language response generation.
- RL Fine-Tuning: Fine-tuned Qwen3-4B-Instruct-2507 with GRPO using the rllm framework to improve strategy adaptation across dialogue turns.
- Reward Design: Designed a composite RL reward integrating XML-format compliance, length control, strategy-following accuracy, and normalized attitude-shift improvement.
- Evaluation: Conducted multi-model evaluation against GPT-4o, Qwen3-4B, Qwen2.5-7B, and LLaMA3.1-8B-Instruct persuadees, and analyzed how persuasion performance changes with conversation length.
Federated Knowledge-Augmented Generation
ZJU-UIUC Institute, China
Spring 2025
Advisor: Qiang Zhang
- Goal: Built a privacy-preserving and personalized text generation framework by combining federated learning with knowledge-augmented generation.
- Method: Designed a federated KAG pipeline that enables distributed training across clients without sharing raw data, while using retrieval-augmented knowledge to improve generation quality.
- Contribution: Collected data, built the project codebase based on LightRAG, and implemented the core training and generation pipeline.
- Link: Github Repo.
ECE445: Senior Design Laboratory
ZJU-UIUC Institute, China
Spring 2025
Member of a four-person team
- Goal: Built a language-guided robotic grasping system that identifies and grasps target objects from spoken instructions.
- Automatic Speech Recognition: Translated spoken instructions into text and extracted the target object label using Wave2Vec 2.0 followed by an NLP model.
- Computer Vision: Used RGB images from the depth camera to detect the target object and estimate its pixel-level bounding box with YOLOE-V8L.
- 6D Pose Estimation: Combined RGB and depth images to estimate the target object's 6D pose, including [x, y, z] coordinates and three Euler angles, using Open3D.
- Motion Planning: Used the estimated 6D pose to choose an appropriate grasping position and map it to motor destination angles.
- PCB Display: Display the name of the target object on a 0.96-inch OLED screen.
- Code and report are available here.
ECE408: CUDA Optimization for LeNet
Urbana, IL
Fall 2024
Individual project
- Streaming: Overlapped data transfer with kernel execution by dividing large vectors into segments and running kernels while copying data between device and host memory.
- Kernel Fusion: Implemented convolution with matrix multiplication using three kernels, then fused them into a single kernel to reduce overhead and improve performance.
- High-Level Libraries: Used Tensor Cores through Warp Matrix Functions and the cuBLAS library.
- Other Optimizations: Applied constant memory for weights, the
__restrict__keyword, and loop unrolling.
ECE391: 391 OS System
Urbana, IL
Fall 2023
Member of a four-person team
- Basic Functionalities: Implemented an operating system supporting scheduling, interrupts, system calls, exceptions, and file systems.
- Self-Designed Features: Added ATA drivers for a writable file system, command history, customizable colors, and autocomplete.
- Systems Programming: Built the kernel and device-driver stack in C and x86, including paging, interrupt handling, terminal support, and course-provided machine problem interfaces.
- Project repository.