Beyond Words: Spectral Imagination for Recurrent Visual Reasoning in Language Models

Abstract

Large language models reason primarily through linguistic representations, but tasks such as spatial simulation, mental rotation, viewpoint transformation, object persistence, and counterfactual scene reasoning can benefit from internal perceptual representations. We propose \method, a recurrent visual-reasoning architecture that couples a pretrained language model with a pretrained latent-diffusion generator and visual encoder, using generated latents as an internal reasoning workspace rather than only as output images. The core contribution is a \emph{multiresolution spectral imagination state}. Diffusion latents are decomposed into smooth frequency bands, while a learned compute-aware controller decides whether imagination is needed and which spectral scales to activate. Separate coarse-to-fine memories persist across reasoning steps, and a learned latent transport operator aligns previous imagined states with current hypotheses before fusion. Each imagined state is re-perceived by the visual encoder, evaluated by a task-utility critic, projected back into the semantic state, and used to determine whether further imagination is required. The framework distinguishes internal critic approval from objective external verification and uses externally verified outcomes for controlled self-distillation. Recursive improvement is implemented through offline candidate generation, verification, retraining, regression testing, and promote-or-rollback. Matched-compute baselines and ablations test whether spectral recurrence provides benefits beyond conventional generated-image reasoning. The accompanying repository includes an executable reasoning loop, multiresolution memory, learned gating and band allocation, latent alignment, pretrained adapters, benchmark manifests, evaluation tools, and regression tests. The work is presented as a methods proposal without fabricated benchmark gains.

Keywords