Deep Review analyzes technical documents.
Google Deep Review is an agentic AI tool that checks long documents for logic, math, facts, and prior work, launched after a large pilot with privacy.
Key Takeaways
- Google Research publicly launched Deep Review, an agentic AI tool that checks documents for logical consistency, mathematical correctness, factual support, and comparison with prior work, following a pilot on more than 16,000 manuscripts.
- The tool addresses a validation bottleneck caused by generative AI increasing submissions and review demand faster than human review capacity can grow.
- Deep Review uses a four-stage pipeline—segmentation, adaptive budgeting, parallel verification, and global synthesis with grounding—to handle long, complex documents while leaving higher-level judgment to humans.
- Public access is tiered: two free reviews per week, higher limits via a Gemini API key, API integration, and initial support for PDF and plain text, with privacy safeguards that exclude uploaded documents from model training.
- The tool extends beyond academic papers to design documents, presentations, legal and financial documents, and introduces a Teamwork Deep Document Review Pattern in Antigravity.
Meta description: Google releases Deep Review publicly after pilots across 16,000 manuscripts, offering agentic AI document verification for researchers, engineers and creators.
Tags: AI, LLM, Google Research, document review, peer review
Featured image: An abstract friendly illustration showing layered documents, linked nodes, and a calm review agent checking claims, with no real people, logos, or text.
Google Research made Deep Review available to the public on October 14, 2026. Rajesh Jayaram and Vincent Cohen-Addad, research scientists at Google Research, described the tool as an agentic AI system that checks complex documents for logical consistency, mathematical statements, factual support and comparison with prior work. The launch follows pilot deployments across more than 16,000 submissions at STOC, ICML, NeurIPS and ICLR, and expands the product from mathematical and scientific peer review to broader knowledge documents, including Google Docs and presentations.
A public launch after a 16,000-manuscript pilot record
Deep Review is the public-facing evolution of the Paper Assistant Tool, an experimental system Google built to help researchers test manuscripts before conference submission. For one year, Google worked with four flagship computer science conferences: STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. Authors could choose to receive private pre-submission feedback on more than 16,000 manuscripts. The feedback was separate from official review decisions. Google said the pilot did not collect paper data for training, and documents were deleted promptly after each program concluded.
The company published further details in the manuscript “Towards Automating Scientific Review with Google’s Paper Assistant Tool.” Author reactions were positive, and Google said the tool improved with model updates and a better harness. The addition of Gemini 4 Argon support increased capabilities beyond earlier versions of the pilot tool.
The release addresses what Google calls a validation bottleneck. Generative AI has made it easier to draft papers, run code, design experiments and write proposals, but human review capacity has not grown at the same pace. This is especially visible in AI conferences. Google reported over 74,000 submissions to ICLR, NeurIPS and ICML this year, up from approximately 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023. Each submission needs multiple high-quality reviews, adding pressure on experts.
Access, limits and privacy
Google opened Deep Review worldwide with a tiered access model. Each user gets two complimentary document reviews per week. Users with high-volume workflows can connect a Gemini API key to obtain higher usage limits and use the Deep Review API for programmatic integration. The first supported input formats are portable document format and plain text; Google said multi-file input is coming soon.
Privacy is a central design point. Google states that documents uploaded to Deep Review will never be used to train, fine-tune, or improve Google models. Uploaded documents are stored for one year only to display them to the user in the Deep Review UI. This makes the tool relevant for users who need automated scrutiny but do not want their unpublished drafts, legal clauses or financial tables absorbed into model training.
How Deep Review checks long documents
A single LLM call can struggle with long, complex documents. It may exhaust its token reasoning budget, miss dependencies across sections, or accept nuanced claims without enough scrutiny. Deep Review uses a four-stage agentic pipeline to spread the work across specialized agents and then reconcile their findings.
- Document segmentation: A segmenter agent breaks the input into logical parts, such as introduction, methodology, experiments or other document sections. Segments may overlap, but together they must cover the full document.
- Adaptive budgeting: The system estimates information density and assigns different compute budgets and specialized inference scaling pipelines to different sections. A dense mathematical proof can receive more review effort than a related-work passage.
- Deep parallel verification: Specialized reasoning agents review segments in parallel. Each agent can see the full document context for cross-references, while focusing on its assigned section. This keeps individual review tasks manageable and coordinates the parallel review swarms.
- Global synthesis and agentic grounding: A synthesis agent merges segment-level reports, checks consistency across sections, reruns logical checks if needed, and performs external grounding checks and tool calls as a final safeguard against hallucinations.
The goal is not to replace human judgment. Deep Review handles some of the cognitively heavy parts of review—checking claims, consistency, evidence and comparisons—while leaving higher-level evaluation to people.
From research papers to everyday documents
Although Deep Review began with academic papers, its methods are broader. The same mechanisms that test whether a proof follows from assumptions can surface gaps in documents that make claims requiring evidence. Google positions the tool for knowledge work beyond science.
For design documents, it can look for unstated architectural assumptions, edge cases and concrete improvement suggestions. For presentations, it can check whether slide transitions flow logically and whether bullet-point takeaways match charts. For legal and financial documents, it can test internal consistency across clauses and data tables, flag ambiguous terms or downstream consequences, and ground claims against supporting evidence.
Google also introduced the Teamwork Deep Document Review Pattern, an implementation pattern inspired by Deep Review and built into Antigravity with the Teamwork multi-agent orchestration framework. The pattern is designed to evaluate research papers, manuscript drafts and technical proposals through a multi-agent hierarchy with self-correction and segment-based rubrics. Antigravity users can invoke it with the /teamwork-preview command.
What the launch means next
Deep Review is presented as a way to widen access to rigorous pre-submission critique. In computer science, detailed review often depends on senior faculty, well-resourced labs or large peer networks. Automated logical verification can help researchers and other knowledge workers stress-test assertions, find avoidable flaws, strengthen documents and enter review processes with more confidence.
Google says it expects tools like Deep Review to play a significant role in peer review over the coming years by helping both authors and reviewers. Next steps include continued improvements through Gemini 4 Argon support and multi-file input. When available, multi-file support should let users run consistency checks across related documents, spreadsheets and exhibits at the same time.
The project is a collaboration across Google Research, Google DeepMind and Google Cloud. Google credited core contributors and research collaborators 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 included Katherine Chou, Corinna Cortes, Burak Gokturk, Vahab Mirrokni, Tomas Pfister and Yossi Matias, along with the program chairs of STOC, ICML, NeurIPS and ICLR.
Labels: General Science, Machine Intelligence, Natural Language Processing

Comments (0)