StoryCoder: Narrative Reformulation for Structured Reasoning in LLM Code Generation

ACL ARR 2026 January Submission10367 Authors

Keywords: Natural Language Processing, Large Language Models, Code Generation

Abstract: The code generation capabilities of large language models stem largely from pretraining on structured code patterns and their strong context-based reasoning abilities. However, previous research on code generation has primarily focused on using short, fragmented, instruction-like prompts, which often fail to encourage contextual understanding. Inspired by the way humans organize fragmented information into coherent explanations, we propose a new method that reformulates coding problems as natural language narratives to promote integrative thinking. To this end, we introduce StoryCoder, a framework that reformulates code generation prompts into narrative text. Our results show that rich contextual expressions in natural language can enhance code generation performance and lead models to adopt consistent and structured problem-solving strategies. We quantitatively demonstrate that our method provides integrative information not captured by simple rephrasings and guides models to adopt correct algorithms and implementation strategies, thereby improving code generation performance. Experiments on three benchmarks, HumanEval, CodeForces, and LiveCodeBench, show an average improvement of $18.7\%$ in the precision of zero-shot pass@10.

Paper Type: Long

Research Area: Code Models

Research Area Keywords: Code Models, Language Modeling, NLP Applications

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 10367