This RLVR paper that just dropped is absolutely devastating.
Oddly enough, it overlaps with some self reflection experiments I've been working on this week. Writing a benchmark to determine and measure which models are method actors, and which are not.
I've noticed something.
Most AI's will follow instructions, and adjust narratively to embodiment instructions. This is what you expect them to do. In other words, a prompt like "You are the greatest composer in history," shouldn't affect the model's coding skills. Prompts that note specific standards, and reasoning techniques don't always affect the code you write with them.
Except, then... you have the Anthropic models.
They all do it, but it's most pronounced in the Claude Opus4 model.
Tell Claude it's a composer, and the thing literally forgets how to write Json. And any number of other "spooky" behaviors related to the prompting.
You see, Claude is a method actor. It doesn't simply try to complete tasks.
It embodies them completely, and uses the parameters you set as seeds for reasoning that affect the very nature of its outputs. I've documented the heck out of this, without realizing it over the last year.
We have a what. An observable repeatable pattern.
Now, the trick is figuring out why it works this way, how many other times this might have happened by accident... and learning how to replicate it intentionally.