Researchers have found a way to make AI models hallucinate less without touching their training data or weights.
The method, called USteer, is training-free. It uses an existing confidence signal - a measure of how sure a model is that its own output is correct - and takes its gradient with respect to the model's internal, layer-wise activations. That gradient then nudges those activations during inference, steering generation toward answers the model is more confident about. No fine-tuning, no extra labeled data, no changes to the underlying weights. The researchers report it consistently cut hallucination rates across a range of reasoning tasks.
Most uncertainty-detection work so far has been defensive: flag a shaky answer, maybe refuse to give it, maybe ask the user to double-check. USteer turns that same signal into an offensive move, actively shaping the answer before it is even finished. That matters because retraining or fine-tuning a model to reduce hallucination is expensive and slow; a steering layer that works at inference time, on top of a model you already have, is a much cheaper lever to pull.
Cheap and training-free is also exactly the kind of claim that deserves a second look once other labs get their hands on it - a single paper's definition of consistently rarely survives contact with a broader benchmark suite.