CV

You can also access the PDF version here: Yuqi Peng’s CV
Latest update: November, 2025

Education

Northeastern University, Sep. 2023 - Apr. 2026 (expected)
- M.S. in Computer Science
- GPA: 4.0 / 4.0
University of Colorado at Denver, Sep. 2019 - May 2024
- B.S. in Mathematics & B.A. in Economics
- GPA: 3.5 / 4.0

Diffusion Models, Vision-Language Models, Model Generalization, Parameter-Efficient Fine-Tuning

Nov. 2024 – Aug. 2025

Generalizable VLM Fine-tuning Methods

Developed GLAD, a generalizable fine-tuning framework for vision-language models in few-shot settings.
Investigated limitations of CLIP prompt tuning and identified causes of overfitting.
Proposed a LoRA-based cross-modal fine-tuning strategy updating only a small portion of parameters.
Designed AlignNet, a lightweight alignment module to enhance text embeddings using image-aware cues.
Introduced SAM-inspired gradient regularization to improve robustness and guide optimization toward flatter minima.
Conducted evaluations on 15 datasets, covering base-to-novel generalization and domain transfer.
Demonstrated that GLAD outperforms CoOp, CoCoOp, MaPLe, and PromptSRC.
Led the full project: conceptualization, method design, implementation, large-scale benchmarking, and writing.

Personalized Image Editing via Diffusion Models

Developed TARA, a framework enabling training-free multi-concept image generation with diffusion models.
Identified token interference and spatial misalignment as key issues in LoRA-based personalization.
Designed Token Focus Masking (TFM) to constrain each LoRA update to its associated rare token.
Proposed Token Alignment Loss (TAL) to enforce alignment between rare-token attention and class-token regions.
Implemented TARA on Stable Diffusion v1.5 & SDXL, enabling multi-concept composition during inference.
Conducted experiments on DreamBooth datasets using CLIP-T, CLIP-I, DINO metrics.
Demonstrated superior identity preservation over DB-LoRA, Mix-of-Show, and other baselines.
Led the full project from algorithm design, implementation, ablation studies, to manuscript writing.

Data-Efficient Subset Selection for Image Classification

Built a data-centric selection system using submodular optimization to choose diverse and informative samples.
Analyzed limits of dataset distillation and random subset selection.
Used ResNet-18 to extract features and guide iterative sample selection.
Evaluated subsets (10%–50%) on ResNet-18/50/101 and ViT.
Achieved 89.16% accuracy at 50% subset size (vs. 82.17% for random), approaching full-data accuracy of 92.14%.

May 2024 – Sept. 2024

Annotated 3D point cloud motion data aligned with RGB video.
Investigated NeRF-based mesh extraction and evaluated performance on internal datasets.