LASER: Low-Rank Activation SVD for Efficient Recursion
For practitioners deploying recursive architectures, this work offers a practical method to reduce memory usage without sacrificing accuracy, though the approach is incremental.
The paper investigates the activation manifold of Tiny Recursive Models (TRMs) during recursive unrolling, finding that activations lie in a low-dimensional subspace. They propose LASER, a dynamic compression framework that achieves ~60% activation memory savings with no statistically significant accuracy degradation.
Recursive architectures such as Tiny Recursive Models (TRMs) perform implicit reasoning through iterative latent computation, yet the geometric structure of these reasoning trajectories remains poorly understood. We investigate the activation manifold of TRMs during recursive unrolling and find that activations occupy an effectively linear, low-dimensional subspace whose principal directions can be tracked dynamically with cheap power iterations. This suggests that weight-sharing concentrates iterative computation along a small number of dominant eigendirections, and we find that this concentration varies sharply across computational sites. We exploit this structure through LASER (Low-Rank Activation SVD for Efficient Recursion), a dynamic compression framework that maintains an evolving low-rank basis via matrix-free subspace tracking with a fidelity-triggered reset mechanism, achieving ${\sim}60\%$ activation memory savings with no statistically significant accuracy degradation. Our analysis raises questions about how recursive architectures allocate representational capacity during implicit reasoning, and whether this concentration can be exploited to improve the efficiency and stability of latent computation.