Search papers, labs, and topics across Lattice.
This paper introduces PLOT, a novel preference learning method for LLM alignment that leverages Optimal Transport (OT) to minimize the distance between model outputs and human preferences at the token level. By framing preference learning as an OT problem, PLOT preserves the original LLM distribution, improving stability and robustness during fine-tuning. Experiments across human values and logic/problem solving benchmarks demonstrate that PLOT consistently improves alignment performance while maintaining fluency and coherence.
Optimal Transport offers a surprisingly effective and theoretically grounded approach to preference learning, outperforming existing methods in aligning LLMs with human values and reasoning abilities.
Preference learning in Large Language Models (LLMs) has advanced significantly, yet existing methods remain limited by modest performance gains, high computational costs, hyperparameter sensitivity, and insufficient modeling of global token-level relationships. We introduce PLOT, which enhances Preference Learning in fine-tuning-based alignment through a token-level loss derived from Optimal Transport. By formulating preference learning as an Optimal Transport Problem, PLOT aligns model outputs with human preferences while preserving the original distribution of LLMs, ensuring stability and robustness. Furthermore, PLOT leverages token embeddings to capture semantic relationships, enabling globally informed optimization. Experiments across two preference categories - Human Values and Logic&Problem Solving - spanning seven subpreferences demonstrate that PLOT consistently improves alignment performance while maintaining fluency and coherence. These results substantiate optimal transport as a principled methodology for preference learning, establishing a theoretically grounded framework that provides new insights for preference learning of LLMs.