Safe, Scalable Transfer with Dynamic Experts: Mitigating Negative Transfer

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Safe, Scalable Transfer with Dynamic Experts: Mitigating Negative 
Transfer"

By

Mr. Zhili LIU


Abstract:

Large-scale pre-training has become a central paradigm in modern machine 
learning, enabling models to acquire broadly reusable knowledge from massive 
and heterogeneous data. However, transfer is not always beneficial: knowledge 
learned during pre-training may interfere with downstream adaptation or give 
rise to unsafe and undesired behaviors. This thesis studies such failures 
through the lens of negative transfer and argues that, in modern pre-training, 
negative transfer manifests in two closely related forms: semantic 
interference in representation learning and implicit harmful knowledge 
transfer in generative models.

To address these challenges, this thesis develops a sequence of methods that 
progressively move from post-hoc mitigation to training-time and architectural 
specialization. First, it studies implicit harmful knowledge transfer in 
diffusion models and proposes Geom-Erasing, a post-hoc framework that 
selectively removes implicitly acquired harmful concepts without retraining 
the entire model. Second, it investigates semantic interference in 
self-supervised representation learning and introduces Scalable Dynamic 
Routing (SDR), a task-customized pre-training framework that reduces harmful 
interference through structural specialization across semantically distinct 
data subsets. Third, it proposes Mixture of Cluster-Conditional Experts 
(MoCE), which extends this idea to a mixture-of-experts framework and improves 
selective knowledge reuse through finer-grained expert allocation. Finally, it 
extends the specialization perspective to self-alignment in large language 
models through Mixture of insighTful Experts (MoTE), showing that structured 
expert allocation and coordinated reasoning can further improve safety, 
robustness, and controllability.

Across diffusion models, self-supervised representation learning, and LLM 
self-alignment, the proposed methods consistently show that structured 
specialization can mitigate harmful or ineffective transfer.

Taken together, these studies suggest that safe and scalable transfer depends 
not only on learning shared representations at scale, but also on structuring 
how knowledge is allocated, specialized, suppressed, and reused under 
heterogeneous downstream requirements.


Date:                   Tuesday, 19 May 2026

Time:                   10:00am - 12:00noon

Venue:                  Room 2128A
                        Lift 19

Chairman:               Dr. Ding PAN (PHYS)

Committee Members:      Prof. James KWOK (Supervisor)
                        Dr. Long CHEN
                        Dr. Dan XU
                        Prof. Can YANG (MATH)
                        Prof. Sinno Jialin PAN (CUHK)