Researchers find a way to read what large language models are actually doing inside their brains — and predict whether they'll get the answer right

Computer scientists developed a new method to analyze what happens inside large language models by studying patterns in how they pay attention to information as they work. Instead of studying individual behaviors in controlled lab settings, this approach scales up to find recurring patterns across millions of inferences, making it possible to predict whether a model will produce correct output without needing to check the answer.

For years, interpretability research has been stuck in a bind: either you get precise explanations of specific model behaviors that don't generalize to real-world use, or you try to study the model at scale and hit a wall of computational expense. This work breaks that deadlock by showing attention patterns are a reliable signal that works across different models and different code tasks. That's significant because it means we can finally measure whether a deployed language model is likely to fail before it fails, and potentially steer it toward correct answers on specific tasks without retraining the whole system.