Metaphors Can Steer LLMs to Inefficient Code

Large language models (LLMs) often benefit from natural language elements like metaphors and analogies, which aid in generalization across different domains. However, this research reveals a downside: metaphorical expressions can inadvertently transfer inappropriate procedural patterns, leading to undesirable behaviors, particularly in code generation. The study shows that metaphorical instructions can steer LLMs towards less efficient algorithms, a phenomenon termed "metaphorical algorithmic steering." This steering occurs when a skill or procedure, benign and plausible in its source domain, transfers an abstract procedural schema into a programming task. This causes the LLM to favor inefficient methods like exhaustive search, full scans, or repeated reconstruction, even without explicit mention of these target algorithms. This suggests that LLMs can carry over procedures appropriate for a task's background domain into a programming context, resulting in suboptimal code. To investigate this, the MASC (Metaphorical Algorithmic Steering for Code Generation) framework was developed. MASC iteratively metaphorizes and refines benign skills to elicit low-efficiency code while maintaining task relevance. Beyond behavioral evaluation, the study found that this phenomenon is detectable and mechanistically reflected in model representations, with metaphorical skills inducing a hidden-state shift towards lower-efficiency procedural behavior prototypes.