AI researchers cut the time to test LLM capability mid-training from an hour to three minutes

Researchers built a lightweight probe that watches the internal patterns of a language model during training and predicts how well it will perform on downstream tasks, cutting evaluation time by 95 percent. This means AI labs can now test whether a model is learning useful skills without waiting an hour to run full evaluations, letting them make faster decisions about when to stop training or adjust approach.

Right now, scaling language models is expensive and slow. Teams train for weeks, then spend hours running tests to see if the result is actually useful. This probe collapses that feedback loop: you can peek at what the model is learning in real time instead of waiting until training finishes. The practical effect is simpler: faster iteration means cheaper experimentation, which means more people can afford to build and refine models, and model development stops being purely a play for organizations with unlimited compute budgets.