Small AI models hallucinate differently depending on how their internal uncertainty patterns shift during decoding

Researchers traced how four tiny language models (1 to 1.7 billion parameters) actually fail when asked to distinguish truth from falsehood, finding that models fall into three categories based on how their internal confidence evolves as they generate text. The finding suggests that you can't fix these models just by tuning the final answer—you have to understand and reshape how they manage uncertainty at every step.

Until now, we've mostly measured whether small models get the right answer or not. This work shows the mechanism underneath: some models get progressively more confident as they generate text (which means they double down on mistakes), others get less confident, and a few stay steady. The implication is practical: if you're deploying a small model on a phone or edge device and you need it to be honest about what it doesn't know, you now have a way to measure which type you're using and what knobs to turn to make it more reliable.