LLMs Fail on Code Semantic Shifts With Operator Redefinitions

JL

Jessy Li@jessyjliTECH

Why? These LLMs aren't conditioning on the provided rules instead they're leaning on pretrained lexical associations. When code gets weird or the semantics shift, the "understanding" breaks down. Such semantic shifts are not contrived or synthetic either and are routinely encountered in mainstream languages that support operator overloading. (3/5)