TRAC: Token-level Reward Assignment for Coherent Abstractive Summarization

摘要

Large Language Models (LLMs) have achieved remarkable success in text summarization, particularly through the integration of reinforcement learning. However, maintaining logical coherence and contextual consistency remains a pervasive challenge in long-form generation, often hindering the production of high-quality, unified summaries. To address these persistent issues, we propose TRAC, a framework that introduces a token-level reward function by integrating relative sentence gain, inter-sentence attention, and a Gaussian length penalty. By training a Process Reward Model (PRM) to provide fine-grained, step-wise supervision, TRAC ensures superior structural integrity and fluency during the generation process. Experimental results demonstrate that TRAC outperforms the sequence-level baseline by 11.05% in Fluency and 10.61% in Relevance. Furthermore, it achieves significant gains over competitive baselines such as FIGA and TLCR, underscoring its effectiveness and generalizability in high-quality NLP summarization.