Efficient Reasoning with Flow Language Models

We study the reasoning efficiency of continuous Flow Language Models (FLMs) relative to discrete diffusion language models. Unlike discrete diffusion, which progressively commits positions to hard tokens, FLM maintains a continuous state and discretizes only after denoising. We theoretically show that this continuous state can represent geometric superpositions of candidate sequences and accumulate their soft evidence along the FLM trajectories. Experiments on graph planning and Sudoku show that, under small denoising-evaluation budgets, FLM achieves higher reasoning accuracy than matched discrete diffusion baselines. Intermediate-state interventions further provide evidence that information beyond the top candidate contributes to subsequent refinement. These results suggest that continuous states provide a useful representation for efficient reasoning with limited denoising evaluations.