面向自动提示词工程的反馈进化算法OA

To address the problems that existing discrete prompt optimization algorithms are overly dependent on large-scale search feedback or limited to natural selection and phrase-level random variation,the feedback-driven evolutionary(FDE)algorithm is proposed,which achieves efficient prompt generation and optimization by integrating active environ-mental feedback and passive natural selection mechanisms.The algorithm combines explicit feedback and implicit guidance,introduces a mutation strategy driven by bidirectional environmental signals,and improves the upper confidence bound evaluation method based on batch sample update,significantly enhancing optimization efficiency.Comparative experi-ments are conducted on 18 natural language task datasets with 8 typical prompt algorithms.Test results show that the com-prehensive performance of the FDE algorithm is relatively improved by 11.44%and the stability by 43.88%compared with the existing best algorithms,while effectively reducing the number of model calls and transmitted characters.Further verification of generalization through cross-parameter scale model analysis reveals that its comprehensive performance is relatively improved by more than 10%compared with the baseline algorithm across all tested models.The FDE algorithm realizes the evolutionary paradigm shift from passive filtering to active-passive collaborative adaptation,which is condu-cive to driving language models to perform natural language processing tasks in a low-cost and more efficient manner.