Evolve Wisely: Decomposing Genetic Algorithms for Chinese Ci Poetry Generation

ACL ARR 2026 January Submission10777 Authors

Keywords: Ci poetry, LLM, Generation, Evolutionary Algorithm

Abstract: Chinese Ci poetry generation requires balancing strict prosodic constraints with aesthetic quality. We decompose Genetic Algorithms (GA) into \textbf{selection} and \textbf{evolution} components, evaluating across 3 LLMs, 8 Cipai formats, and 48 prompts (144 configurations, 5,760 poems). Our findings reveal that GA is highly effective: multi-sampling with selection achieves +19.2\% improvement over zero-shot. Decomposition analysis shows that selection accounts for 96.3\% of performance gain, serving as a powerful ``drafting'' phase. The ``revision'' phase (evolution) is where methods diverge. LLM-guided operators achieve 40\%+ success rates on capable models (DeepSeek: 43.8\%, GPT-5.1: 41.7\%), significantly outperforming the blind mutation baseline (character-level: $\sim$25\%, $p < 0.01$). While mechanical operators actively degrade quality, semantic-aware evolution successfully simulates a human poet's intelligent revision process. Our results suggest a strategy to \textbf{"evolve wisely"}: rely on selection for robust baselines, and reserve expensive semantic evolution for capable models to achieve quality breakthroughs.

Paper Type: Long

Research Area: Natural Language Generation

Research Area Keywords: Generation; analysis; automatic evaluation; inference methods; text-to-text generation;

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: Chinese

Submission Number: 10777