Controlling Recurrent AI Dynamics for Deeper Problem Solving

Recurrent neural networks (RNNs) are often designed with the hope that iterating their update longer at test time will enable them to solve more complex problems, a concept known as "test-time depth." However, additional iterations can unpredictably improve, maintain, or even degrade the quality of an answer. This paper reveals that a measurable property of the trained operator, its "finite-time dynamical regime" (categorized as settling, marginal, or drifting), directly indicates which outcome will occur. The researchers establish a sufficient condition for "depth-safety": once the operator's per-step displacement becomes small relative to the decoder margin, further iterations will not alter the decoded answer. Empirically, they demonstrate that "settling" operators do not degrade with added depth and can even improve accuracy on harder, unseen instances of algorithmic tasks, such as Sudoku. A key finding is that a single terminal fixed-point objective can effectively control this dynamical regime and, consequently, the depth behavior. Removing this objective can induce drifting and eliminate performance gains, while adding it to a generic recurrence yields depth-safe extrapolation. The paper provides four operational criteria for useful test-time depth, catalogues common failure modes, and applies these measurements to existing models like Huginn-3.5B, classifying them within the non-settling family.