Align to Structure: Aligning Large Language Models with Structural Information

Generating long, coherent text remains a challenge for large language models (LLMs), as they lack hierarchical planning and structured organization in discourse generation. We introduce Structural Alignment, a novel method that aligns LLMs with human-like discourse structures to enhance long-form text generation. By integrating linguistically grounded discourse frameworks into reinforcement learning, our approach guides models to produce coherent and well-organized outputs. We employ a dense reward scheme within a Proximal Policy Optimization framework, assigning fine-grained, token-level rewards based on the discourse distinctiveness relative to human writing. Two complementary reward models are evaluated: the first improves readability by scoring surface-level textual features to provide explicit structuring, while the second reinforces deeper coherence and rhetorical sophistication by analyzing global discourse patterns through hierarchical discourse motifs, outperforming both standard and RLHF-enhanced models in tasks such as essay generation and long-document summarization.

Authors

• Zae Myung Kim University of Minnesota
• Anand Ramachandran Amazon
• Farideh Tavazoee Amazon
• Joo-Kyung Kim Amazon
• Oleg Rokhlenko
• Dongyeop Kang University of Minnesota

Keywords

No keywords are indexed for this paper.