Dual-Dimensional Consistency: Balancing Budget and Quality in Adaptive Inference-Time Scaling

Problem
This paper addresses the inefficiencies in inference-time scaling of Large Language Models (LLMs), particularly the trade-off between sampling budget and reasoning quality. Existing methods often treat the dimensions of sampling width and depth as independent, leading to suboptimal performance. Width consensus methods can exacerbate hallucinations, while depth pruning can prematurely cut off valid reasoning paths. The authors propose a novel approach, Dual-Dimensional Consistency (DDC), to unify these dimensions and improve the quality of reasoning while optimizing resource usage. This work is presented as a preprint and has not yet undergone peer review.

Method
The core technical contribution is the Dual-Dimensional Consistency (DDC) framework, which integrates two key components: a Confidence-Weighted Bayesian protocol and a Trend-Aware Stratified Pruning mechanism. The Confidence-Weighted Bayesian protocol assesses the reliability of reasoning paths, allowing the model to focus computational resources on those paths that exhibit high confidence. The Trend-Aware Stratified Pruning selectively terminates less promising paths based on their historical performance trends, thereby enhancing the overall reasoning quality. The authors do not disclose specific details regarding the architecture of the LLMs used, nor do they provide explicit training compute metrics, but they evaluate DDC across five distinct benchmarks.

Results
The DDC framework demonstrates significant improvements in efficiency and accuracy. Specifically, it achieves over a 10x reduction in token consumption compared to strong baseline methods while maintaining or exceeding their accuracy across various LLMs. The paper reports that DDC outperforms existing state-of-the-art methods on all five benchmarks tested, although specific numerical results and comparisons to named baselines are not detailed in the abstract.

Limitations
The authors acknowledge that while DDC effectively balances budget and quality, it may still be sensitive to the initial configuration of the Bayesian protocol and the pruning strategy. They do not address potential limitations related to the scalability of the method across different model architectures or the generalizability of the results to other tasks outside the evaluated benchmarks. Additionally, the reliance on historical performance trends may introduce biases if the model encounters novel reasoning scenarios.

Why it matters
The implications of this work are significant for the field of adaptive inference in LLMs. By providing a unified framework that effectively manages the trade-off between sampling budget and reasoning quality, DDC could lead to more efficient deployment of LLMs in resource-constrained environments. This approach not only enhances the practical usability of LLMs but also opens avenues for further research into adaptive inference techniques that can dynamically adjust to varying computational budgets while maintaining high-quality outputs. The findings could influence future work on model optimization and the development of more robust reasoning mechanisms in LLMs.

Authors: Rongman Xu, Yifei Li, Tianzhe Zhao, Yanrui Wu, Bo Li, Hang Yan
Source: arXiv:2605.15100
URL: https://arxiv.org/abs/2605.15100v1