Invariant-Measure Reasoners: Stable Representations for Latent Reasoning
Invariant-Measure Reasoners (ImR) address prediction instability in latent reasoning models where states converge to a compact set instead of a fixed point. ImR uses an invariant measure, which describes the long-run distribution of these states, to predict from the expectation of the prediction head's output under this measure. This stabilizes predictions and improves accuracy on tasks like maze and Sudoku.