Direct Pass@k optimization balances quality and diversity for efficient test-time reasoning

Framing prefix thought pruning as a submodular pass@k objective allows reasoning models to discard unpromising rollouts early while preventing redundant candidate clustering.

Primary source: arXiv:2610.31667: https://arxiv.org/abs/2610.31667

What’s new: Test-time scaling enables reasoning models to solve complex mathematical and logical problems by generating multiple candidate rollouts and selecting the best outcome. Because generating hundreds of long chain-of-thought solutions to completion demands massive computational budgets, recent methods prune candidate pools after generating short prefixes. However, existing pruning techniques divide into two conflicting heuristics: quality-based pruning selects partial solutions with high estimated success scores, which frequently clusters similar trajectories that fail on identical edge cases, while diversity-based pruning enforces embedding dispersion, often retaining fundamentally flawed reasoning paths. In Direct Pass@k Optimization for Thought Pruning: Balancing Quality and Diversity in Test-Time Reasoning (arXiv:2610.31667, 22 Oct 2026), researchers Mahdi Salmani, Adel Javanmard, Devansh Gupta, and Meisam Razaviyayn from the University of Southern California formulate prefix thought pruning as a direct optimization of the pass@k metric.

Rather than relying on separate heuristics, the authors formalize thought pruning by maximizing the joint probability that at least one surviving trajectory in the pruned subset of size k successfully resolves the problem when completed. By modeling the conditional success probability and inter-rollout correlation, the pass@k objective naturally acts as a submodular function: it favors high-quality individual rollouts while penalizing mutual redundancy. Evaluated across eight open-weight reasoning models on rigorous reasoning benchmarks, direct pass@k pruning consistently outperforms both pure score-based filtering and embedding exploration. The efficiency advantages are especially noticeable in moderate-budget regimes, where continuing only a small subset of candidate prefixes matches the accuracy of full-compute parallel rollouts while cutting inference FLOPs substantially.

Why it matters: Scaling inference-time compute has become the dominant paradigm for frontier problem-solving models, but brute-force sampling exhausts inference clusters quickly. Test-time pruning offers a principled path to high-accuracy reasoning under fixed inference budgets. By directly optimizing the target metric (pass@k) instead of relying on disconnected proxy scores, this approach establishes a rigorous mathematical foundation for early stopping and tree search in reasoning architectures.

Caveats: Estimating pairwise trajectory correlations and continuation probabilities from short prefix tokens requires training calibrated value probes or lightweight scoring estimators, which adds training complexity. Furthermore, on problems where early prefix tokens are nearly indistinguishable between correct and incorrect reasoning paths, prefix-based pruning risks discarding viable solutions before critical reasoning steps unfold.