Search papers, labs, and topics across Lattice.
This paper investigates the dynamics of recurrent-depth reasoners, revealing that the finite-time dynamical regime of the trained operator鈥攚hether it is settling, marginal, or drifting鈥攄etermines the impact of additional iterations on answer quality. The authors establish a condition for depth-safety, showing that when the per-step displacement is small relative to the decoder margin, further iterations do not degrade the decoded answer. Empirical results demonstrate that settling operators can enhance accuracy on unseen instances, such as improving Sudoku performance from 0.19 to 0.34, while also providing operational criteria to identify failure modes in depth extrapolation.
Settling operators can significantly boost accuracy on unseen tasks, transforming additional iterations into a performance advantage rather than a risk of degradation.
Recurrent-depth reasoners aim to solve harder problems by iterating their update longer at test time, but additional iterations can improve, preserve, or degrade an answer. We show that a measurable property of the trained operator, its finite-time dynamical regime (estimated as settling, marginal, or drifting), indicates which of these occurs. We give a sufficient condition for depth-safety: once an operator's per-step displacement is small relative to the decoder margin, the decoded answer cannot change under further iterations. Empirically, on algorithmic tasks trained from $800$ unaugmented examples per difficulty tier, settling operators do not degrade with added depth, and on some tasks convert it into higher accuracy on harder unseen instances (Sudoku, $0.19$ to $0.34$ past the training horizon). A single terminal fixed-point objective moves the regime and the depth behavior together: removing it induces drift and removes the gains, and adding it to a generic recurrence yields depth-safe extrapolation on carry propagation. We give four operational criteria for useful test-time depth, use them to catalogue failure modes, and, as a consistency check, apply the same measurements to Huginn-3.5B, which falls in the non-settling family.
