Abstract
Large language models (LLMs) learn their behavior from training data, yet tracing a generation back to the data responsible for it is prohibitively expensive at LLM scale. This dissertation develops a data-centric approach to scalable and trustworthy LLMs, centered on RapidIn, which estimates the influence of each training example on a given generation at the sequence and token level. RapidIn compresses gradient vectors by over 200,000× for caching and achieves a 6,326× speedup in retrieval, estimating the influence of an entire dataset on any generation within minutes. Data influence then guides decisions about data and models. ALinFiK introduces the Linearized Future Influence Kernel for prospective data valuation and distills it into a lightweight model, significantly reducing GPU memory, storage, and time. RASLIK applies randomized antipodal search on linearized influence kernels to select forget and retain data for unlearning, reducing variance and lowering retrieval time and memory. RapidUn uses RapidIn's token-wise influence estimates to weight parameter-efficient unlearning updates, achieving up to 100× speedup over full retraining and consistently improving forgetting–retention trade-offs over gradient ascent, Fisher, and LoReUn. At inference time, UniGuardian treats prompt injection, backdoor, and adversarial attacks as one category, Prompt Trigger Attacks, and detects them with a training-free mechanism whose Single-Forward Strategy runs detection and generation in a shared forward pass. Finally, DMin extends scalable influence estimation to diffusion models with billions of parameters, reducing storage from around 40 GB to 80 KB per sample and retrieving the most influential training samples with k-nearest-neighbor search. Together, these studies show that compact, reusable gradient representations make data influence practical at scale and useful for data valuation, selective model updates, and malicious-input detection.
Publication Date
10-2026
Document Type
Dissertation
Student Type
Graduate
Degree Name
Computing and Information Sciences (Ph.D.)
Department, Program, or Center
Computing and Information Sciences Ph.D, Department of
College
Golisano College of Computing and Information Sciences
Advisor
Weijie Zhao
Advisor/Committee Member
Zhiqiang Tao
Advisor/Committee Member
Yidan Hu
Recommended Citation
Lin, Huawei, "Scalable and Trustworthy Large Language Models: A Data-Centric Approach" (2026). Thesis. Rochester Institute of Technology. Accessed from
https://repository.rit.edu/theses/12836
Campus
RIT – Main Campus
