Google launches Deep Review to verify complex documents

Google has officially released Deep Review, a public agentic AI system designed to provide rigorous verification and detailed feedback on complex technical text. Evolving from an experimental initiative called the Paper Assistant Tool, the platform was previously trialed across four major computer science conferences—STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027—where authors submitted more than 16,000 manuscripts for pre-submission feedback. The launch addresses a severe validation bottleneck caused by a surge in research writing; submissions across ICML, NeurIPS, and ICLR spiked from approximately 23,000 in 2023, 32,000 in 2024, and 45,000 in 2025 to over 74,000 in 2026.

Unlike standard single-prompt language models that struggle with long-context reasoning, Deep Review utilizes a four-stage pipeline. A segmenter agent breaks inputs into logical sections, followed by an adaptive budgeting step that dynamically scales inference compute for dense segments such as mathematical proofs. Specialized reasoning agents then conduct parallel verification across individual sections before a synthesis agent consolidates the reports, double-checks internal logic, and uses external tools to prevent hallucinations. Powered by Gemini 4 Argon, the system extends beyond academic manuscripts to scrutinize architectural design documents, slide decks, legal agreements, and financial reports.

Every user gets two complimentary document reviews per week, with programmatic integration and higher volume limits accessible by connecting a Gemini API key. Input initially supports PDFs and plain text, with multi-file capabilities scheduled for future updates. Google specifies that uploaded documents are stored for one year solely to display results in the UI and are never used to train or fine-tune models. Developers using Antigravity can also trigger the Teamwork Deep Document Review Pattern via the /teamwork-preview command.

For researchers and technical professionals, Deep Review democratizes access to high-level pre-submission critique, allowing authors to stress-test logical assertions and catch architectural edge cases prior to formal review or publishing.