Google Research launches Deep Review, an agentic AI system for technical document verification

Google Research has publicly released Deep Review, an agentic AI system designed to provide rigorous feedback and technical verification on complex documents. Evolving from the experimental Paper Assistant Tool (PAT), which processed over 16,000 submissions across STOC, ICML, NeurIPS, and ICLR, the system now addresses the validation bottleneck created by the rapid increase in AI-generated content.

Four-Stage Agentic Pipeline

Deep Review overcomes the limitations of single-prompt LLM calls through a specialized four-stage architecture:

• Document segmentation: An agent decomposes the input into logical, overlapping segments to ensure full coverage.
• Adaptive budgeting: The system dynamically allocates compute resources based on information density, giving more depth to complex sections like mathematical proofs.
• Deep parallel verification: Specialized reasoning agents scrutinize segments in parallel, tracking cross-references while focusing on specific review tasks.
• Global synthesis and agentic grounding: A synthesis agent consolidates reports, performs consistency checks, and executes external grounding checks to prevent hallucinations.

Expanded Capabilities and Access

While initially focused on computer science academic papers, Deep Review now supports a broader range of knowledge documents, including Google Docs, presentations, legal agreements, and financial documents. The system checks logical consistency, grounds facts, and flags ambiguous terms.

• Availability: Users receive two complimentary document reviews per week. Higher limits are available via the Gemini API key.
• Privacy: Uploaded documents are never used to train or fine-tune Google models and are stored for one year solely for user display.
• Integration: A Teamwork Deep Document Review Pattern is now available in Antigravity, allowing users to invoke multi-agent review workflows via the /teamwork-preview command.

The release supports PDFs and plain text, with multi-file input support planned for the future. The underlying model has been updated with Gemini 4 Argon support to enhance analysis quality.