Topology-Enhanced Alignment for Large Language Models: Trajectory Topology Loss and Topological Preference Optimization
Year: 2026
Type: conference-paper
Abstract: Alignment of large language models (LLMs) typically relies on supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF), or more recently direct preference optimization (DPO).However, existing objectives largely ignore the global geometry and topology of the representation space: they operate on local token-level likelihoods or scalar preference scores, and do not explicitly constrain how hidden states move from a user prompt to an answer. * Equal contribution.
Authors: Yurui Pan, Ke Xu, Bo Peng
Institutions: Fudan University, Tongji University, Shanghai Ocean University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166879139