Variation in Verification: Understanding Verification Dynamics in Large Language Models
Summary: arXiv:2509.17995v2 Announce Type: replace-cross
Abstract: Recent advances have shown that scaling test-time computation enables large language models (LLMs) to solve increasingly complex problems across diverse domains. One effective paradigm for test-time scaling (TTS) involves LLM generators producing multiple solution candidates, with LLM verifiers assessing the correctness of these candidates without reference answers. In this paper, we study generative verifiers, which perform verification by generating chain-of-thought (CoT) reasoning followed by a binary verdict.
We systematically analyze verification dynamics across three dimensions – problem difficulty, generator capability, and verifier generation capability – with empirical studies on 12 benchmarks across mathematical reasoning, knowledge, and natural language reasoning tasks using 14 open-source models (2B to 72B parameter range) and GPT-4o.
Our experiments reveal three key findings about verification effectiveness:
- Easy problems: Easy problems allow verifiers to more reliably certify correct responses.
- Weak generators: Weak generators produce errors that are easier to detect than strong generators.
- Verification ability: Verification ability is generally correlated with the verifier’s own problem-solving capability, but this relationship varies with problem difficulty.
These findings reveal opportunities to optimize basic verification strategies in TTS applications. First, given the same verifier, some weak generators can nearly match stronger ones in post-verification TTS performance. For example, the performance gap between Gemma2-9B and Gemma2-27B shrinks by 75.7%. This suggests that the capability of the generator plays a crucial role in determining the overall effectiveness of the verification process.
Second, we identify cases where strong verifiers offer limited advantage over weak ones. In certain scenarios, both types of verifiers fail to provide meaningful verification gains, indicating that scaling up the verifier alone cannot overcome fundamental verification challenges. This highlights the importance of a balanced approach in developing both generators and verifiers to enhance overall performance.
In conclusion, our study provides valuable insights into the verification dynamics within large language models. As LLMs continue to evolve and are applied to increasingly complex tasks, understanding the interplay between generative capabilities and verification strategies will be essential for optimizing their performance. The findings from our empirical studies pave the way for future research aimed at refining test-time scaling methodologies and improving the reliability of AI systems across diverse applications.