Google has publicly released Deep Review, an agentic system for checking technical documents, after testing its earlier Paper Assistant Tool with more than 16,000 opt-in submissions to STOC, ICML, NeurIPS and ICLR. The conference feedback was separate from official peer review.
The system divides a document into overlapping logical sections, assigns more computation to dense material such as mathematical proofs, and runs specialized reviewing agents in parallel. A final agent combines their findings, checks cross-section consistency and uses external sources to ground claims. Google positions the tool for research papers as well as design, legal, financial and presentation documents, while leaving higher-level judgments to people.
Deep Review accepts PDF and plain-text files, with multi-file analysis planned. Every user receives two complimentary reviews per week. People who need higher limits or API access can connect a Gemini API key. Google says uploaded documents are not used to train or improve its models, although the public interface stores them for one year so users can view them.
The release responds to a growing review bottleneck: Google counts more than 74,000 submissions to ICLR, NeurIPS and ICML in 2026, up from about 45,000 in 2025. Prior conference pilots measured author reactions, not whether automated feedback can replace expert peer review.