That’s Deprecated! Understanding, Detecting, and Steering Knowledge Conflicts in Language Models for Code Generation
A study of how language models handle conflicts between learned programming knowledge and prompt-provided updates, with probes and activation steering for code generation.
Language models for code often face prompts that conflict with what they learned during training: APIs change, libraries deprecate methods, and documentation updates rewrite earlier conventions.
This project studies those conflicts in code generation. It looks at when models follow the prompt versus their prior knowledge, whether those conflicts can be detected from activations, and whether steering can shift the model’s preference.