PETS: A Principled Framework Towards Optimal Trajectory Allocation for Efficient Test-Time Self-Consistency

Researchers introduce PETS, a framework for optimizing how many reasoning trajectories to sample from AI models during inference to maintain accuracy while reducing computational costs. By modeling trajectory allocation as a crowdsourcing problem, the approach achieves up to 75% budget savings on benchmarks while maintaining perfect consistency, addressing a key efficiency challenge in test-time scaling.

Test-time scaling—running multiple reasoning paths and aggregating results—has emerged as a powerful technique for improving AI model outputs, but it comes with significant computational costs. PETS addresses this efficiency problem by introducing a principled optimization framework centered on a new metric called self-consistency rate, which measures agreement with an infinite-budget majority vote. This theoretical grounding distinguishes the work from heuristic approaches and enables rigorous analysis of budget allocation strategies.

The innovation lies in bridging two traditionally separate fields. By modeling reasoning traces as workers in a crowdsourcing problem, the authors leverage decades of established theory from that domain, yielding both theoretical guarantees and practical algorithms. This connection is conceptually elegant and computationally tractable. The framework operates in two complementary modes: an offline setting where all questions are known upfront, and an online streaming setting where questions arrive sequentially and the system must adapt dynamically.

For developers and researchers building AI systems, PETS offers tangible benefits. Achieving 75% budget reduction on challenging benchmarks like GPQA while maintaining perfect consistency translates directly to lower inference costs—a critical consideration for production deployments. The online setting is particularly relevant for real-world applications where questions arrive unpredictably and resource constraints are strict.

Looking forward, this work could influence how AI systems are deployed at scale, particularly in domains where computational budgets are limited but accuracy cannot be compromised. The open-source code availability enables rapid adoption and validation across different model architectures and datasets. Future research may extend these allocation strategies to other test-time scaling paradigms beyond simple majority voting.

• →PETS achieves up to 75% reduction in sampling budget while maintaining perfect test-time consistency on GPQA through principled trajectory allocation.
• →The framework models AI reasoning traces as crowdsourcing workers, leveraging established theory to provide theoretical guarantees and efficient algorithms.
• →Self-consistency rate, a new metric measuring agreement with infinite-budget majority votes, provides a theoretically grounded measure for optimization.
• →Dual offline and online settings enable both advance planning and dynamic adaptation to question difficulty in streaming scenarios.
• →Open-source implementation enables practical adoption for researchers and developers seeking to reduce inference costs in production systems.