Google opens Deep Review, an AI agent that checks technical documents, after testing it on 16,000+ papers

Google Research announced on October 14 the public release of Deep Review, which it describes as an agentic AI system engineered to provide rigorous feedback and technical verification on complex documents. In a post by research scientists Rajesh Jayaram and Vincent Cohen-Addad, Google said the system was initially designed to analyze computer science papers and has since expanded to documents in domains such as legal and finance.

Deep Review is the evolution of Google's experimental Paper Assistant Tool (PAT). Google says that over the past year it partnered with four computer science conferences, STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027, to give private, pre-submission feedback to the authors of more than 16,000 manuscripts who opted in. The feedback was decoupled from the official review process, the data was not collected for training, and papers were deleted after each program, according to the post. Google says this year's three major AI conferences drew over 74,000 submissions, up from roughly 45,000 in 2025.

Google describes a four-stage pipeline: an agent splits the document into logical segments; compute is budgeted by how dense each section is; specialized reasoning agents scrutinize segments in parallel with access to the full document; and a synthesis agent consolidates the reports, reruns consistency checks and performs external grounding checks and tool calls as a guard against hallucinations. The tool now supports Gemini 4 Argon.

Every user gets two complimentary document reviews per week, and heavier users can connect a Gemini API key for higher limits and programmatic access through a Deep Review API. It initially accepts PDFs and plain text, with multi-file input coming soon. Google says uploaded documents are never used to train, fine-tune or improve its models and are stored for one year only for display in the Deep Review interface. It is also adding a Deep Review-inspired document review pattern to its Antigravity coding tool.