Tag
This paper investigates how language models execute in-context conditional rules by probing whether testing and routing are separable mechanisms. Using activation patching across three open models and six languages, the authors find that predicate testing is modular while route representations are token-bound and non-transferable.