Budget-aware Test-time Scaling via Discriminative Verification
Kyle Montgomery, Sijun Tan, Yuqi Chen, Siyuan Zhuang, Tianjun Zhang, Raluca Ada Popa, Chenguang Wang
2025-10-17
Summary
This paper explores ways to improve the accuracy of large language models, specifically when they're tackling difficult problems that require careful reasoning.
What's the problem?
Currently, a popular method to boost accuracy involves having another AI 'verify' the initial answer generated by the first AI. However, this verification step, when done by a generative AI, is very computationally expensive, meaning it requires a lot of processing power and time, making it impractical for many real-world uses.
What's the solution?
The researchers investigated a different type of verification – discriminative verification – which is less demanding on computing resources. They found that while discriminative verification isn't great on its own, combining it with a technique called 'self-consistency' (where the AI generates multiple answers and picks the most common one) creates a surprisingly effective and efficient system. This combined approach actually outperformed the more expensive generative verification method, achieving better results on a challenging reasoning benchmark.
Why it matters?
This research demonstrates that you can significantly improve the performance of large language models without needing massive amounts of computing power. It offers a practical and cost-effective way to make these powerful AI tools more accessible and useful for everyday applications, showing that a 'smarter' approach to verification can be better than simply throwing more resources at the problem.
Abstract
Test-time scaling is a powerful strategy for boosting the performance of large language models on complex reasoning tasks. While state-of-the-art approaches often employ generative verifiers to select the best solution from a pool of candidates, this method incurs prohibitive computational costs, limiting its practicality. In this work, we shift the focus to a more budget-aware paradigm: discriminative verification. We conduct a thorough empirical analysis and demonstrate that while discriminative verifiers may underperform in isolation, combining them with self-consistency in a hybrid approach creates a powerful and efficient test-time scaling mechanism. Notably, under a fixed compute budget, this hybrid approach surpasses state-of-the-art generative verification by a significant margin: achieving up to 15.3\% higher accuracy on AIME2025. Our findings establish that for practical, real-world applications, budget-aware scaling with discriminative verifiers is not only a "free" upgrade over self-consistency, but also a more effective and efficient alternative to costly generative techniques. Code is available at https://github.com/wang-research-lab/verification.