Title: FineVerify: Enhancing Agentic Search via Fine-Grained Self-Verification and Test-Time Compute Scaling

Original: arXiv:2606.00660v1 Announce Type: new Abstract: Agentic search requires language model agents to explore many sources and answer complex information-seeking questions. Scaling test-time compute is a promising way to improve these agents, but current approaches can fail, because correct answers are often sparse and score-based selection depends on model calibration. We propose FineVerify, a fine-grained self-verification framework that decomposes each question into checkable sub-questions, verifies sampled candidates against each sub-question, and selects the candidate with the highest aggregated score. This per-check structure turns selection into simpler local judgments and produces scores under the same explicit criteria. Across four agentic search benchmarks and two models, FineVerify consistently outperforms standard scaling baselines. With only four sampled trajectories, it improves GPT-5-mini by 8.2 accuracy points and Gemini-3-flash by 5.6% on average. With 12 samples, FineVerify enables GPT-5-mini to surpass frontier GPT-5 on BrowseComp-Plus. Beyond accuracy, FineVerify produces interpretable verification traces that help audit benchmark errors, suggesting broader applications for inspecting agentic search systems. Code and data are available at https://github.com/XuZhao0/fineverify

Rewrite:

Abstract: To address complex information-seeking queries, agentic search systems rely on language model agents that must navigate and evaluate numerous data sources. While increasing test-time computational resources offers a viable path to enhancing agent performance, existing methods frequently falter. This instability arises because correct solutions are often rare, and reliance on score-based selection mechanisms is heavily contingent upon the model’s calibration capabilities.

To overcome these limitations, we introduce FineVerify, a framework for fine-grained self-verification. This approach breaks down complex queries into verifiable sub-questions. It then evaluates sampled candidate responses against each specific sub-question, ultimately identifying the candidate that achieves the highest cumulative score. By structuring the verification process around individual checks, FineVerify transforms the selection task into a series of manageable local judgments, ensuring that scores are generated based on consistent, explicit criteria.

Our evaluation across two distinct models and four agentic search benchmarks demonstrates that FineVerify consistently surpasses standard scaling baselines. Specifically, utilizing just four sampled trajectories, the method boosts the accuracy of GPT-5-mini by 8.2 points and enhances Gemini-3-flash by an average of 5.6%. Furthermore, when employing 12 samples, FineVerify allows GPT-5-mini to outperform the state-of-the-art GPT-5 on the BrowseComp-Plus benchmark.

In addition to improving accuracy, FineVerify generates interpretable verification traces. These traces facilitate the auditing of errors within benchmarks and indicate potential wider utility for inspecting and understanding agentic search systems. The associated code and datasets can be accessed at https://github.com/XuZhao0/fineverify

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC