Energy-Regularized Sequential Model Editing on Hyperspheres

Qingyuan Liu*, Jia-Chen Gu*, Yunzhi Yao , Hong Wang , Nanyun Peng

• Developed SPHERE (Sparse Projection for Hyperspherical Energy Regularized Editing), projecting new knowledge onto sparse subspaces of edited weight to preserve hyperspherical uniformity with rigorous proof linking hyperspherical uniformity and editing stability. Achieved +16.4% average editing capability over baselines while most faithfully preserving general performance on LLaMA3 (8B) and Qwen2.5 (7B).

LAVID: An Agentic LVLM Framework for Diffusion-Generated Video Detection

Qingyuan Liu, Yun-Yun Tsai, Ruijian Zha, Pengyuan Shi, Victoria Li Chengzhi Mao and Junfeng Yang

• Designed an agentic framework LAVID that leverage Large Vision Language Models (LVLMs) in detecting AI-generated videos. The LVLMs will call external tools to extract additional information from the video to facilitate themselves in the detection.
• Evaluation results show LAVID improves F1 score by 6.2% to 30.2% over the top baseline on the comprehensive dataset across four SOTA LVLMs.

Turns Out I’m Not Real: Towards Robust Detection of AI-Generated Videos

Qingyuan Liu, Pengyuan Shi, Yun-Yun Tsai, Chengzhi Mao, and Junfeng Yang

• Developed Diffusion Reconstruction Error (DIRE) based method DIVID to detect AI-generate videos; achieved a detection accuracy of up to 93.7% for videos from their benchmark dataset of videos generated from Stable Vision Diffusion, Sora, Pika, and Gen-2.
• This work was marked as the Columbia Engineering Research Highlight.

BALANCED LATENT SPACE OF DIFFUSION MODELS FOR COUNTERFACTUAL GENERATION

Baohua Yan, Qingyuan Liu, Zhaobin Mo, Kangrui Ruan and Xuan Di

• we propose a framework that balances the latent space by incorporating signals that facilitate the transition to new counterfactuals while preserving factual information. We first identify the cause of this imbalance as the uncontrolled signal from the counterfactuals. Based on this understanding, we introduce a balancing method within the diffusion process.

Causal Adjacency Learning for Spatiotemporal Prediction Over Graphs

Zhaobin Mo*, Qingyuan Liu*, Baohua Yan, Longxiang Zhang, and Xuan Di

• Designed the Causal Adjacency Learning (CAL) framework, enhancing model prediction performance on the ODD dataset; applied with the heuristic method which contained the correlation calculation and Condition Independence Test.
• Achieved 14.23%, 15.49%, and 50.27% RMSE improvement in SafeGraph dataset, compared with the results based on distance, correlation, and attention matrix separately.

Bubble Bobble Game
Report | Presentation

Course Project of CSEE4840 Embedded Systems, CU (Rated 1st)

• Designed and implemented bubble bobble game on an embedded system utilizing both the ARM CPU and the FPGA on the DE1-SoC, with a focus on efficient video and audio processing. The system is based on line buffer, integrating various hardware inputs of VGA, Game Controller, Audio Jack, and Keyboard.

Research on 5G Network Slicing System and Strategy for End Users

Report | Presentation

National Second Prize (Top 3% out of 1006 teams worldwide) in C4

• Applied Reinforcement Learning method to optimize network slicing strategy; tested strategy with 12 cellphones, increasing network throughput by 69.4% and document transfer efficiency by 82.2%.

Microprocessor without Interlocked Pipeline Stages (MIPS) CPU Design

Report

• Designed a CPU from the scratch with MIPS framework on the Logisim platform; Practical structures have been integrated here such as Pipeline Stalling and Brach History table.