acceptodds
Desk rejected at ICLR 2027

Temporal Abstraction and Precision Control through Latent Dynamics

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Temporal abstraction is usually modeled with explicit options, skills, or hierarchical policies. We ask whether option-like temporal abstraction can instead be realized by recurrent latent dynamics without explicit option identities, option-specific policies, or learned termination functions. We study an input-driven recurrent action-value model in which normalized task inputs drive a shared latent dynamical system with learned timescales, while a prediction-error correction pathway provides a learned, context-dependent route for residual signals to update the latent state. We define option-like latent dynamics by operational criteria: persistence beyond one primitive step, behavioral influence on action values, input-dependent initiation, modulation or interruption, reuse, and absence of explicit option labels. In a controlled Four Rooms navigation setting, we test these criteria using impulse responses, latent perturbations, pathway-forcing analyses, precision-demand manipulations, and canonical-frame reuse under known symmetry. The central test is whether effective duration and correction change when temporal cost or observation reliability changes. This framing treats temporal abstraction as a measurable property of latent dynamics rather than as a discrete architectural object. The resulting evidence supports a scoped mechanistic account: recurrent systems can initiate, sustain, correct, and reuse option-like latent dynamics without selecting from an explicit library of options.

Primary area:
reinforcement learning
Keywords:
RL, dynamical-systems, neuroscience, active-inference, precision

Related papers

Discussion (0)

Sign in to comment.