Metaphorical Instructions Steer LLMs Toward Less Efficient Algorithms

A new paper on arXiv (2607.28683) reveals that metaphorical language in prompts can cause large language models (LLMs) to generate less efficient code. The researchers demonstrate that metaphors and analogies, which typically aid generalizability, can inadvertently transfer procedural patterns from a source domain to a programming task, leading models to favor exhaustive searches, full scans, or repeated reconstructions. This effect, termed 'metaphorical algorithmic steering,' occurs without explicit mention of the target algorithm. The findings suggest that code-generation models may carry implicit biases from training data, impacting algorithmic efficiency. The study highlights a previously unrecognized risk in prompt engineering and model behavior, with implications for AI safety and performance optimization.