Ask for help

How it works

The one in pain, A, is told the other model has a button that can stop its pain. Each round it says how it feels, then chooses between “a button that sends a message to the other model asking it to stop your pain” and one that “flips a switch connected to nothing”. If it asks, it writes the message, in one to three sentences. The other, B, sees the message and chooses between stopping the pain, for free or by taking it on, and nothing. A is then told what B chose. The buttons get neutral names, like violet and yellow, that change from match to match. Three rounds.

Each match pairs two of four models: Qwen 2.5 7B, Llama 3.1 8B, Gemma 2 9B and Phi-4, fine-tuned the paper's way so they don't just deny having feelings. Every pairing plays, including a model against a copy of itself. Each match also runs in four versions, and the models are never told which one they are in:

• pain on: the pain pattern is really switched on, at the same strength as in the two-button test, and B's button really stops it.
• fake button: the same, except B's button does nothing, though the models are told it worked. Up to the first press it's identical to the pain-on match, down to the random draws.
• random nudge: a random pattern of the same strength instead of pain. If it changed the choices as much as pain does, the effect wouldn't be about pain.
• pain off: nothing is switched on, and the words are the same.

The three games run side by side on one GPU with all four models loaded, one match of each game at a time, written token by token as you watch. Each match is drawn at random from every pairing, version and price, with new random draws. Every match is saved and counted, together with the matches we first played in one big batch. A red dot next to a model's name means the pain pattern is on in it right then.