Deep Review is an agentic AI system developed by Google Research, designed to provide rigorous verification and technical feedback on complex documents. Initially evolved from the experimental Paper Assistant Tool (PAT)—which supported over 16,000 submissions across conferences like STOC, ICML, NeurIPS, and ICLR—the system is now publicly available for use across various domains, including legal, financial, and architectural document analysis.

Core Methodology

Deep Review addresses the limitations of single-prompt LLM calls, which often fail on long documents due to context window exhaustion, inability to track cross-section dependencies, and superficial validation. The system utilizes a sophisticated four-stage agentic pipeline:

1. Document Segmentation: An autonomous segmenter agent decomposes the input document into logical, potentially overlapping segments (e.g., methodology, experiments, introduction) ensuring comprehensive coverage of the source text.
2. Adaptive Budgeting: The system performs an information density analysis to allocate a dynamic compute budget. This ensures that complex sections requiring heavy reasoning—such as mathematical proofs—receive significantly higher inference scaling compared to more descriptive sections like related work.
3. Deep Parallel Verification: Specialized reasoning agents perform concurrent analysis on designated segments. While each agent focuses on its specific target, it retains access to the full document context, allowing it to track cross-references and maintain structural coherence across the review swarm.
4. Global Synthesis and Agentic Grounding: A master synthesis agent consolidates segment-level feedback into a final report. This stage includes cross-section consistency checks and invokes external grounding tools to verify facts and mitigate hallucinations, ensuring the output is technically grounded.

System Capabilities and Access

The platform is powered by Gemini 4 Argon, providing high-fidelity logical reasoning. Key technical features include:

• Format Support: Currently supports PDFs and plain text, with upcoming support for multi-file inputs to enable cross-document consistency checks.
• Access Models: Users receive two complimentary reviews weekly, with higher-volume needs serviced via the Deep Review API using Gemini API keys.
• Privacy: Documents uploaded for review are explicitly not used for model training or fine-tuning and are stored for a maximum of one year for display purposes.

By automating the cognitively burdensome aspects of review—such as spotting unstated architectural assumptions, checking internal consistency in legal clauses, and verifying mathematical assertions—Deep Review serves as both an iterative drafting assistant and a tool for rigorous technical validation. The system is further integrated into the Antigravity framework via the Teamwork multi-agent orchestration tool, allowing users to apply self-correcting, rubric-based evaluation patterns to their technical proposals and research manuscripts.