StructPO: A New Framework for AI-Generated Paper Introductions

A new framework called StructPO has been developed by researchers to enhance the quality of introductions in academic papers generated by large language models (LLMs). This struct-aware policy learning framework integrates the multi-stage writing process into a single-pass policy, utilizing explicit stage tokens to tackle the difficulties of creating a cohesive narrative that includes background, gap identification, methodology, and contributions. Unlike existing methods that rely on costly multi-stage prompts or agent workflows, which can suffer from cross-stage drift, StructPO employs struct-aware credit assignment to separate local stage quality from overall coherence and incorporates refinement-guided optimization for first-pass policy revision. Experimental results indicate that StructPO outperforms workflow-based baselines in terms of semantic alignment, structural rationality, and inference efficiency, while also generalizing to out-of-domain contexts and competing effectively with GPT-4. The paper can be accessed on arXiv under the identifier 2608.03138.