The steering works by importing a procedural schema, not by naming an algorithm. A metaphor about searching a crowd carries 'check everything one by one' into the code, so the model writes an exhaustive scan without the prompt ever mentioning search strategy.
The authors' MASC framework repeatedly rewrites benign skills into metaphors until the model emits slow code, and the prompts stay plausible and on-task throughout. A content filter looking for malicious phrasing would pass every one of them.
The effect is not surface wordplay. Metaphorical prompts push the model's hidden states toward low-efficiency behavior prototypes, which means the bias lives in the transferred procedure and is detectable in the model's internals.
Why it matters
If you rely on LLMs to generate code, this finding shows that innocuous metaphorical phrasing in your prompts (e.g. 'search the crowd', 'scan the shelf') can silently bias the model toward less efficient algorithms — worth knowing before you trust prompt wording as neutral.
Key quotes
“We refer to this metaphor-induced effect as metaphorical algorithmic steering: a skill that is benign and plausible within its source domain transfers an abstract procedural schema into a programming task, causing the model to favor exhaustive search, full scans, or repeated reconstruction without explicitly mentioning the target algorithm.”
“We also find that metaphorical skills induce a hidden-state shift towards lower-efficiency procedural behavior prototypes.”
articleHow well LLM-based test generation techniques perform with newer LLM versions?Michael Konstantinou, Renzo Degiovanni, Mike Papadakis
articleHow Reasoning Shapes Social Bias in LLM-Generated Code?Weifeng Sun, Jieke Shi, Zhou Yang, Yuchen Chen, Hongyan Li, Meng Yan, David Lo
articleWhat do LLMs think when you don't tell them what to think about?Together AI