Wang, Y. (CSE) – Controlling Undesirable Knowledge and Behaviors in Large Language Models: Unlearning, Evaluation, and Feedback Dynamics

Large language models (LLMs) can acquire and reproduce information and behaviors that later become undesirable, such as private or copyrighted content, hazardous knowledge, and social bias. Controlling them is important but difficult in practice.
LLM unlearning aims to remove the influence of such content from a model while preserving its overall capability. Retraining LLMs from scratch is often prohibitively expensive, motivating practical interventions that may need to operate without retain data or access to model weights.
This dissertation studies how to unlearn or intervene selectively under practical constraints, and what evidence is needed to assess whether an intervention works. It makes four main contributions.
First, we propose FLAT, a loss adjustment method that unlearns using only forget data to control the knowledge.
Second, for models whose weights cannot be updated, we study training-free unlearning at inference time, from embedding-corrupted prompts (ECO) to DRAGON, which detects forget-worthy prompts and guards black-box LLMs through in-context chain-of-thought reasoning.
Third, we introduce UnlearnBench, a clean and controllable benchmark with explicit knowledge provenance, and show that the evaluated methods fail at entity-level unlearning even when retraining succeeds, that multi-stage memorization changes unlearning behavior, and that reasoning can recover supposedly forgotten information.
Finally, we extend our investigation from interventions on individual models to the evolution of undesirable behaviors across successive model generations. We therefore study how social bias evolves when models are repeatedly trained on their own outputs under performative feedback and propose reward-guided selection with reweighting to mitigate bias.
Ultimately, this dissertation advances the understanding of how to control undesirable knowledge and behaviors in LLMs, through methods designed for practical constraints and evaluations that test whether unlearning succeeds, contributing to the responsible development and deployment of LLMs.
Event Host: Yaxuan Wang, PhD Candidate, Computer Science & Engineering
Zoom: https://ucsc.zoom.us/j/7186149767?pwd=y129sEV0UQmywrma4wQO5ffbBG7rnm.1&omn=98767260141
Passcode: 625461