Parallel Speaker–Listener Large Language Models (PaSLLMs): Non-Sequential Text Generation via Token-Set Prediction and Differentiable Reordering

This work presents a theoretical framework for reformulating large language model generation using a parallel Speaker–Listener architecture. The approach decouples semantic content prediction from linguistic ordering through an unordered token-set representation and differentiable reordering mechanisms. The framework aims to explore alternatives to sequential autoregressive decoding and expand the design space for modular, distributed inference in large-scale language models. No empirical results are claimed; the contribution is conceptual and architectural.