Agentic Verifiers: Provably Safe Test-time scaling for Reasoning Models

This project introduces a novel architecture for agentic AI systems that ensures accuracy, efficiency, and safety during reasoning. It addresses two key challenges—lack of steerability and absence of verifiable guarantees—by developing verifiers that can interject at any point in a model’s generation process. An auxiliary monitor model evaluates each reasoning step against predefined properties, rolling back and correcting errors in real time. The research spans commonsense, medical, and legal reasoning, aiming to deliver publicly available verifier models and an open-source platform for integrating verification into agentic systems, paving the way for trustworthy AI in high-stakes domains.

This research is conducted via The Agentic AI Research and Innovation (AARI) Initiative which focuses on the next frontier of agentic systems through Grand Challenges with the academic community and Microsoft Research.