FineVerify: Scaling Test-Time Compute with Fine-Grained Self-Verification for Agentic Search

Researchers introduced FineVerify, a self-verification framework for agentic search that breaks complex questions into smaller checkable sub-questions.

It scores sampled agent trajectories against these local checks and selects the highest-scoring candidate, making test-time verification more granular and interpretable.

On four agentic search benchmarks with two models, FineVerify beat standard scaling baselines, including gains for GPT-5-mini and Gemini-3-flash.

With 12 samples, it even helped GPT-5-mini surpass GPT-5 on BrowseComp-Plus.