Parallel-Probe: Towards Efficient Parallel Thinking via 2D Probing

TL;DR

Parallel-Probe introduces a novel 2D probing interface and a training-free controller to enhance the efficiency of parallel reasoning in large models, reducing computational costs while maintaining accuracy.

Contribution

It presents 2D probing to reveal global dynamics in parallel thinking and a controller that optimizes reasoning depth and width without additional training.

Findings

Reduces token cost by up to 35.8% compared to majority voting.

Achieves over 25.8% reduction in total token cost.

Establishes a superior Pareto frontier for test-time scaling.

Abstract

Parallel thinking has emerged as a promising paradigm for reasoning, yet it imposes significant computational burdens. Existing efficiency methods primarily rely on local, per-trajectory signals and lack principled mechanisms to exploit global dynamics across parallel branches. We introduce 2D probing, an interface that exposes the width-depth dynamics of parallel thinking by periodically eliciting intermediate answers from all branches. Our analysis reveals three key insights: non-monotonic scaling across width-depth allocations, heterogeneous reasoning branch lengths, and early stabilization of global consensus. Guided by these insights, we introduce $\textbf{{Parallel-Probe}}$, a training-free controller designed to optimize online parallel thinking. Parallel-Probe employs consensus-based early stopping to regulate reasoning depth and deviation-based branch pruning to dynamically adjust width. Extensive experiments across three benchmarks and multiple models demonstrate that Parallel-Probe establishes a superior Pareto frontier for test-time scaling. Compared to standard majority voting, it reduces sequential tokens by up to $\textbf{35.8}$% and total token cost by over $\textbf{25.8}$% while maintaining competitive accuracy.