A new arXiv preprint asks how language models handle contradictory evidence inside a prompt. The research centers on lexical redefinition: an explicit instruction tells the model to interpret a word in a particular way, and then a subsequent sentence appears to conflict with that definition. This setup is designed to reveal whether models follow the instruction or get swayed by the surrounding context.

The abstract, as released, introduces the research question and the experimental approach but cuts off mid-sentence. Because the full text is not yet available, the article does not report specific results or conclusions. What is clear is the paper's focus on context-sensitive instruction following, a key issue for building reliable and controllable language models.

Readers interested in the full methodology and findings will need to consult the complete paper on arXiv. The current abstract alone establishes the problem and the proposed testing framework, leaving the outcome of the study open until the full version is read.