Funnel of Thoughts: Efficient Test-Time Scaling via Early Voting and Rollout Pruning

A new arXiv paper (2608.15065) introduces Funnel of Thoughts (FoT), an inference-time method for Large Reasoning Models (LRMs) that reduces computational cost while preserving accuracy. The method addresses the high expense of multi-sample inference, which is necessary for reliable deployment due to inconsistent answers across repeated queries. Majority voting at k rollouts is the standard but costly approach. FoT maintains the full 32-trajectory voted accuracy while halving attention FLOPs, achieving a 28.8% reduction in full-model inference cost. The method leverages a training-free lexical signal: unproductive trajectories often contain hesitation markers like "Wait", "Actually", and "perhaps". These trajectories are less likely to reach correct answers and consume disproportionate attention FLOPs, sometimes degenerating into no-answer loops. The paper analyzes 115K reasoning trajectories from six LRMs. FoT uses early voting and rollout pruning to improve efficiency. The paper is available on arXiv under the announcement type "new". The authors are not named in the provided content. The method is significant for making LRMs more practical for real-world applications by reducing inference costs.