Google has announced the public release of Deep Review, an agentic, multi-agent AI system designed to deliver rigorous technical feedback and verification for long, complex documents. Originally created to analyze computer science research papers, Deep Review’s scope now includes legal, financial and other domain-specific documents. It can be used both during drafting for iterative improvement and as an aid in the formal review process.
Why this matters
Generative AI has greatly increased the pace of technical content creation—papers, code, experimental designs and proposals are being produced more rapidly than before. Human verification capacity, however, has not scaled at the same rate, producing a validation bottleneck: information is generated faster than it can be thoroughly reviewed. This is especially visible in academic peer review, where submissions to major AI conferences have risen sharply: this year ICLR, NeurIPS and ICML received over 74,000 submissions combined, up from roughly 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023. Each submission requires multiple high-quality reviews, increasing the burden on human experts.
From PAT pilots to a public product
Deep Review evolved from Google’s experimental Paper Assistant Tool (PAT). Over the past year Google partnered with four flagship computer science conferences—STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027—to provide private, pre-submission feedback to more than 16,000 authors who opted in. The pilots aimed to give actionable feedback ahead of submission; this feedback was separate from official conference reviews, PAT did not use the papers to train models, and submissions were deleted from Google systems after each program ended. With Gemini 4 Argon support added to Deep Review, Google states the tool’s underlying capabilities have improved relative to earlier PAT iterations.
How Deep Review operates
Single-prompt large language model calls can fail on lengthy, complex documents by exhausting token budgets, missing long-range cross-sectional dependencies, or failing to scrutinize nuanced claims. Deep Review addresses these problems with a four-stage agentic pipeline:
- Document segmentation: a segmenter agent decomposes the document into logical sections (e.g., Introduction, Methods, Experiments). Segments may overlap and must cover the entire document.
- Adaptive budgeting: the pipeline inspects information density and dynamically assigns compute budgets so that dense sections—such as mathematical proofs—receive more compute than less demanding sections.
- Deep parallel verification: specialized reasoning agents analyze each segment in parallel; each agent can access full document context but focuses its compute on its assigned section to enable manageable, coordinated review swarms.
- Global synthesis and agentic grounding: a synthesis agent consolidates segment reports, runs cross-section consistency checks, and performs external grounding checks and tool calls as a final measure against hallucinations.
Beyond research papers: wider use cases
Although Deep Review began with academic paper analysis, the same long-context reasoning mechanisms generalize to other document types where claims must be supported by evidence. Google highlights several use cases:
- Design documents: identify unstated architectural assumptions, surface edge cases, and suggest concrete improvements.
- Presentations: verify logical slide flow and that bullet-point takeaways match visual charts.
- Legal and financial documents: check internal consistency across clauses and tables, flag ambiguous terms and downstream consequences, and ground claims against supporting evidence.
Access, formats and data policy
- Tiered access and API: every user receives two complimentary document reviews per week. Users with high-volume workflows can connect their Gemini API key for higher usage limits and programmatic integration via the Deep Review API.
- Input formats: initial support includes PDFs and plain text; multi-file input support is coming soon.
- Privacy and data policy: documents uploaded to Deep Review are not used to train or fine-tune Google models. Uploaded documents are stored for one year solely for display in the Deep Review UI.
Antigravity and the Teamwork review pattern
Google is also introducing the Teamwork Deep Document Review Pattern, a Deep Review–inspired implementation built into Antigravity using the Teamwork multi-agent orchestration framework. This pattern applies a multi-agent hierarchy with self-correction and segment-based rubrics to evaluate research papers, drafts and technical proposals. Antigravity users can invoke it via the /teamwork-preview command.
Expected impact and next steps
Comprehensive pre-submission critique in computer science has traditionally depended on access to extensive review networks—senior faculty, well-resourced labs or broad peer advisory circles. Automated, reliable logical verification could help authors stress-test claims, find avoidable flaws, strengthen arguments on verified foundations, and enter the review process with greater confidence. Google plans to continue expanding Deep Review’s capabilities: Gemini 4 Argon support has improved feedback quality, and upcoming multi-file support will enable consistency checks across related documents, spreadsheets and exhibits.
Getting started
- Try Deep Review: users can submit documents via the Google Research Portal.
- Integrate with the API: developers can incorporate programmatic verification into CI and drafting workflows using a Gemini API key.
- Read the research: a preprint describes the methodology and pilot evaluations.
Acknowledgments
Deep Review is a collaboration across Google Research, Google DeepMind and Google Cloud. Core contributors include Santiago Balseiro, Taylan Cemgil, Tommaso D'Orsi, Kirk Draheim, Lindi Emoungu, Ayse Erduran, Palash Goyal, Caroline Hartnell, Shobana Kannan, Amy Keeling, Siyi Kou, Bridget Lee, Ian Li, Anoushka Mazumdar, Taylor Montgomery, Mouad Naciri, Uche Okonkwo, Mihir Parmar, Anthony Phalen, Rory Pilgrim, Sarah Presotto Post, Rob Tierney, Drew Tyler, Vishnu V., Vahab Mirrokni, Laura Vardoulakis, Johannes Welbl, David Woodruff, Yunhan Xu and Jinsung Yoon. Executive sponsors named are Katherine Chou, Corinna Cortes, Burak Gokturk, Vahab Mirrokni, Tomas Pfister and Yossi Matias; Google also thanks the Program Chairs of STOC, ICML, NeurIPS and ICLR for their partnership.
