A new pruning method aims to improve the efficiency of reasoning models by directly optimizing how candidate solutions are selected for further computation. Developed by researchers at the University of Southern California, the technique, detailed in a paper on arXiv, addresses a core challenge in test-time scaling: balancing the generation of high-quality answers with the need for diverse reasoning paths.
Test-time scaling involves generating numerous potential solutions to a problem, with the expectation that at least one will be correct. However, running every candidate solution to completion is computationally expensive. Thought pruning offers a solution by cutting off unpromising reasoning paths early, after only a short prefix has been generated. Current pruning strategies generally fall into two categories: quality-based methods, which score partial solutions and discard those likely to be incorrect, and diversity-based methods, which select paths that explore different reasoning trajectories to cover a wider range of potential outcomes. The tension between these two approaches mirrors the exploration versus exploitation dilemma in machine learning.
The new method, termed Direct Pass@k Optimization for Thought Pruning (PKTP), directly optimizes the pass@k metric. This metric measures the probability that at least one of k generated solutions is correct. By maximizing this objective, PKTP inherently balances the selection of high-quality partial solutions with the penalization of redundant ones. This unified approach means that quality and diversity are balanced by the criterion itself, rather than through separate, ad-hoc design choices.
Experiments conducted on eight open-weight reasoning models show that PKTP outperforms existing quality-based and diversity-based pruning strategies. The improvements are particularly notable when a moderate number of partial solutions are retained for continued generation, underscoring the value of directly balancing quality and diversity during the pruning process. This work challenges the notion that pruning inherently degrades test-time scaling performance, suggesting that carefully designed pruning methods can maintain or even enhance it.
The researchers model the joint distribution of candidate solution correctness using a latent Gaussian model. In this model, the correctness of each candidate is determined by thresholding a Gaussian score. This approach allows for the optimization of pass@k, which depends on the joint distribution of candidate outcomes, as similar partial solutions tend to succeed or fail together. The method's effectiveness has been demonstrated across various reasoning models and sizes, showing consistent superiority in both moderate and large pruning scenarios without manual tuning.