Researchers discover code LLMs can be steered toward specific programming languages and libraries by manipulating activation vectors during inference.

Scientists used difference-in-means methods to estimate layer-wise steering vectors for five language/library pairs across three open-weight code LLMs

Activation space interventions substantially increased code generation toward target ecosystems even when prompts explicitly requested different choices

Steering effectiveness varied by model and target, with common ecosystems easier to induce than rarer alternatives

Code-style preferences appear to be represented as compact, steerable structures in activation space rather than being fixed in model parameters