Deep Review was announced by Google Research on October 14, 2026 as publicly available worldwide. It is an agentic system built to review long documents: it does not write the draft for you, it checks whether the draft's logic holds together, whether the mathematics is sound, and whether its claims are backed by evidence.
Google's framing is blunt: content generation keeps getting faster while the human capacity to verify it stays flat. Submissions to just three conferences, ICLR, NeurIPS, and ICML, passed 74,000 this year, up from roughly 45,000 in 2025. Deep Review grew out of an experimental tool called the Paper Assistant Tool, which over the past year partnered with four flagship conferences, STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027, to give pre-submission feedback to authors of more than 16,000 manuscripts who opted in. The pilots were fully separate from official peer review, and the manuscripts were not used for training.
How the four agent stages divide the work
The system splits reviewing into four stages. First, segmentation: the document is decomposed into overlapping logical segments so the full text is covered. Second, adaptive budgeting: information-dense sections, such as mathematical proofs, receive a larger reasoning budget, while introductions and related work receive less. Third, deep parallel verification: specialized agents each scrutinize one segment while retaining access to the full document context, so cross-references can be tracked. Fourth, global synthesis and grounding: a synthesis agent merges the segment reports, runs cross-section consistency checks, and performs external grounding checks to guard against hallucination. The public release also adds Gemini 4 Argon support, lifting capability beyond the pilot versions, and the same approach ships as a document review pattern inside Antigravity, invoked with /teamwork-preview.
On access, every user gets two complimentary reviews per week, and high-volume users can connect their own Gemini API key for programmatic use through the API. Initial input formats are PDF and plain text, with multi-file input coming soon. The privacy terms are explicit: uploaded documents are never used to train or improve Google models, and they are stored for one year solely so users can view them in the interface. The scope goes beyond papers: design documents, presentations, and legal and financial documents are all listed, and the system hunts the same class of problems everywhere, namely internal contradictions, ambiguous wording, and data that does not match the conclusions. One judgment should be kept: Google itself stresses that higher-level calls remain with humans. Deep Review works well as a first filter before submission or signing; treating it as the review verdict itself would go too far.
