Google releases Deep Review for automated technical document verification

Google Research has publicly released Deep Review, an agentic artificial intelligence system designed to provide rigorous technical verification for complex documents. Following successful private pilots with over 16,000 submissions at major computer science conferences including STOC, ICML, NeurIPS, and ICLR, the tool is now available to a wider audience of researchers and engineers. The system aims to address the growing imbalance between the rapid generation of AI-assisted content and the fixed human capacity for detailed peer review.

What happened

The volume of technical submissions to academic venues has surged dramatically, creating a bottleneck in the peer review process. Submissions to three major AI conferences (ICLR, NeurIPS, and ICML) rose from approximately 23,000 in 2023 to over 74,000 in 2026. Each manuscript requires multiple high-quality reviews to maintain scientific integrity, placing an unsustainable burden on human experts who possess the necessary specialized knowledge. Deep Review was developed to alleviate this pressure by automating the cognitively demanding aspects of verification, such as checking logical consistency and grounding factual claims.

Previously known as the Paper Assistant Tool (PAT), the system operated as an experimental initiative where authors could opt-in for pre-submission feedback. This feedback was strictly decoupled from the official conference review processes. During these pilots, which included partnerships with STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027, the tool helped authors stress-test their manuscripts. Google reports that author reactions were extremely positive, citing improvements in both the underlying models and the evaluation harness. With this public launch, the system has been rebranded as Deep Review and expanded beyond academic papers to support domains like law and finance.

The public release introduces tiered access and new integration capabilities. Every user receives two complimentary document reviews per week. For high-volume workflows, users can connect their own Gemini API key to access higher usage limits and programmatic integration via the Deep Review API. Currently, the system supports PDFs and plain text inputs, with multi-file support scheduled for a future update. Google explicitly states that documents uploaded to Deep Review are never used to train or fine-tune its models and are stored for only one year for display purposes in the user interface.

How it works

Deep Review avoids the limitations of single-prompt large language model calls, which often fail on long documents due to token budget exhaustion or overlooked cross-section dependencies. Instead, it employs a four-stage agentic pipeline. First, a segmenter agent ingests the document and decomposes it into logical segments, such as the introduction, methodology, or experiments. These segments may overlap to ensure full coverage of the content. Second, an adaptive budgeting mechanism inspects the information density of each section. It dynamically allocates compute resources, assigning more inference scaling to dense mathematical proofs than to lighter sections like related work.

In the third stage, specialized reasoning agents scrutinize each segment in parallel. While each agent focuses on its specific target section, it retains access to the full document context to track cross-references and maintain coherence. This parallel swarm approach ensures that reviewing tasks remain manageable while maintaining coordination. Finally, a dedicated synthesis agent consolidates the segment-level reports. It performs global consistency checks between sections and reruns logical checks if necessary. This agent also conducts external grounding checks and tool calls to prevent hallucinations and verify facts against existing knowledge.

Key details

• Deep Review has processed over 16,000 submissions during pilot programs at STOC, ICML, NeurIPS, and ICLR.
• The system uses a four-stage agentic pipeline: segmentation, adaptive budgeting, parallel verification, and global synthesis.
• Users receive two free document reviews per week, with option to use a personal Gemini API key for higher limits.
• Supported input formats currently include PDFs and plain text, with multi-file support coming soon.
• Uploaded documents are not used for model training and are deleted after one year.
• The tool is powered by Gemini 4 Argon, which enhances its capability to handle complex logical and mathematical verification.

Why it matters

For software engineers and technical founders, Deep Review represents a shift toward automated rigor in document-heavy workflows. As AI accelerates the creation of code, design documents, and technical proposals, the ability to quickly verify logical consistency becomes critical. The system’s ability to catch unstated architectural assumptions, spot edge cases, and suggest concrete improvements can significantly reduce the time spent on manual code reviews and design audits. By offloading the tedious aspects of verification, teams can focus on higher-level strategic decisions and innovation.

Furthermore, the expansion of Deep Review beyond academia signals a broader applicability for enterprise use cases. Legal and financial documents often suffer from internal inconsistencies and ambiguous terms that can lead to costly errors. A tool that can systematically check clauses against data tables and ground claims in evidence offers a powerful safeguard for compliance and risk management. The availability of an API allows organizations to integrate this verification layer directly into their continuous integration and drafting workflows, ensuring that technical documentation remains accurate and reliable throughout the development lifecycle.

What you can do

• Visit the Google Research Portal to try Deep Review with your first two complimentary document reviews.
• Integrate the Deep Review API into your drafting workflow by connecting your Gemini API key for programmatic access.
• Use the tool to stress-test technical design documents for logical gaps and unstated assumptions before implementation.
• Apply Deep Review to legal or financial contracts to check for internal consistency and ambiguous terminology.
• Read the published preprint to understand the underlying methodology and evaluation metrics used in the pilot studies.
• Monitor updates for multi-file support to eventually run consistency checks across entire sets of related documents and spreadsheets.