Viewpoint Invariance

A new Meta paper argues that likelihood is the wrong test of truth for LLMs, since models learn what is common in text rather than what is true, making frequently repeated falsehoods seem probable. Drawing on physics and philosophy, it treats a fact as real only if it survives a change of viewpoint: the model answers from several perspectives (e.g., three medical specialists), keeps what all answers share, and abstains when nothing is shared. The paper runs no experiments and relies on worked examples, such as a binary case where a loss of 0.67 yields 100% accuracy while a nearly equal loss of 0.71 yields 0%.