Advised by Prof. Hongkai Xiong and Prof. Wenrui Dai, my researches include:

• Proposed modular Chebyshev connections into general network layers to improve the approximation capability of a neural network. Specifically, established recursive relationship among adjacent layers and polynomial relationship between non-adjacent layers. Comprehensive experiments vefify its strong approximation and optimization capability. (ICME 2025 on submission)
• Proposed Residual Side Tuning (RST) framework to enhance information extraction efficiency by employing a dual-block side-tuning structure with low-rank linear mapping on aggregated features, and introducing element-wise feature enhancement strategy to model cross-layer information. The properties and performance of RST is verified through mathematical proof and various experiments. (ICML 2025 on submission)
• Proposed GLEAM, an efficient fine-tuning method for large model parameters. This method leverages the high similarity of parameter matrices in LoRA to construct a low-rank decomposition, further reducing the number of parameters required for fine-tuning while enhancing performance. (ICME 2025 on submission)
• Proposed PAT, a fault-tolerant multimodal classifer to solve the problem of conflict between prompter learning and adapter tuning. Specifically, it aligns model representations obtained from Soft Prompt and Adapter-based methods and incorporats contrastive learning loss to enhance model performance and generalization. (ICML 2025 on submission)

Advised by Prof. Jieping Ye, my researches include:

• Proposed SalaMAnder, a Shapley-value-based framework for quantifying component-level contributions in CoT reasoning. Specifically, we develop an efficient stratified sampling algorithm to compute Shapley value for mathematical expression attribution and CoSP (Cardinality of Shapley Positives) metric. Theoretical derivation and comprehensive validation across multiple models and benchmarks present a robust monotonic correlation between CoSP and model performance, providing theoretical explanations for the empirical success of CoT. (ACL 2025 on submission)
• Proposed CRFT, a novel method that identifies and optimizes critical representations that integrate significant information from preceding layers or regulate subsequent layer representation. CRFT effectively optimizes the representations in a low-rank linear subspace through information flow analysis. (ACL 2025 on submission)
• Classfied and summarized research work on interpretability on large language models.

Advised by Prof. Quanshi Zhang, my researches include:

• Theoretically derived the analytical solution for multi-step adversarial attacks, which explains the reasons behind the optimization difficulties in adversarial training. This is validated through experimental results. (Accepted by AAAI 2024)
• Theoretically derived the two-stage dynamic interaction process of DNNs, proving that the network learning process gradually encodes interactions of varying complexity. This provides a theoretical foundation for understanding overfitting. (Accepted by NeurIPS 2024)
• Theoretically derived and validated the robustness of concepts with different complexities.

Advised by Prof. Chongyang Zhang, my researches include:

• Developped Swin Transformer based model to implemente instance segmentation of workpiece welding area.
• Designed a space-time filter to remove false positive samples in pedestrian detection.
• Developped YOLOv5-based model to detect tower crane, recognize dangerous tower crane, and label electronic fence.

My research and engineering projects include:

• Developed various models to complete intraday prediction tasks, surpassing the baseline model in global domains (A-shares) and improving performance in live trading within three weeks. Conducted ablation studies on the model to explain the reasons for performance improvement and validate its generalizability.
• Fine-tuned the above model to enhance performance by 10% in local domains (ZZ800 and ZZ1000) while maintaining minimal decrease in overall performance.