In a research preprint published on arXiv on October 22, 2026, computer scientists Mahdi Salmani, Adel Javanmard, Devansh Gupta, and Meisam Razaviyayn from the University of Southern California (USC) introduced a principled algorithmic framework for test-time thought pruning in reasoning models. Titled 'Direct Pass@k Optimization for Thought Pruning: Balancing Quality and Diversity in Test-Time Reasoning', the paper targets the primary computational bottleneck of test-time scaling: the practice of generating hundreds of parallel reasoning chains in the hope that at least one trajectory reaches a valid answer.
While running dozens of extensive chain-of-thought solutions to completion delivers substantial accuracy gains on mathematics and competitive programming benchmarks, it imposes prohibitive inference costs. To curb compute budgets, modern serving engines generate short reasoning prefixes, score the candidates early, and prune unpromising trajectories before committing further tokens. However, prior pruning techniques divided sharply into two disjoint camps: quality-based strategies that evaluate individual prefixes via process reward models, and diversity-based approaches that cluster prefixes to maximize solution variety. The resulting tension represents a classic exploration-versus-exploitation dilemma that existing runtimes manage through fragile, ad-hoc heuristics.
The USC team resolved this structural limitation by formulating an optimization objective that maximizes pass@k directly over the surviving candidate set. Rather than ranking partial thoughts in isolation, the algorithm selects a subset of prefixes specifically calculated to maximize the joint probability that at least one chosen trajectory, when rolled out to completion, will produce an accurate answer. Because the pass@k objective inherently penalizes redundant candidate solutions that arrive at the same failure modes or identical reasoning steps, the criterion naturally balances trajectory quality and portfolio diversity in a unified mathematical formulation.
To validate the approach, the researchers conducted extensive empirical evaluations across eight distinct open-weight reasoning architectures. The direct pass@k formulation consistently outperformed both isolated quality scoring and dedicated diversity clustering methods. The performance advantage was especially pronounced in realistic deployment environments where inference budgets restrict runtimes to retaining a moderate number of surviving prefixes. The framework provides a rigorous foundation for next-generation inference engines seeking to squeeze maximum reasoning reliability out of constrained compute clusters.