LASER: Low-Rank Activation SVD for Efficient Recursion

TL;DR

This paper introduces LASER, a dynamic low-rank compression method for recursive models that significantly reduces activation memory usage without sacrificing accuracy by exploiting the low-dimensional structure of activations.

Contribution

LASER leverages the low-rank structure of activation manifolds in recursive models to enable efficient, dynamic compression with minimal accuracy loss.

Findings

Activations in TRMs occupy a low-dimensional subspace that can be tracked dynamically.

LASER achieves approximately 60% reduction in activation memory.

Weight-sharing concentrates computation along a few dominant eigendirections.

Abstract

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.