On October 14, 2026, Google Research announced the public launch of Deep Review — an agentic AI system that thoroughly reviews complex documents and provides technical feedback on them. The tool grew out of academic peer review and now works with legal, financial, and engineering texts, presentations, and Google Docs documents.
Where Deep Review came from
Deep Review is an evolution of the experimental Paper Assistant Tool (PAT). Over the past year, Google tested it with four leading computer science conferences: STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027. Authors of more than 16,000 manuscripts voluntarily received private feedback from the system before submitting their papers. This feedback had no effect on official peer review, and the papers themselves were deleted after the programs ended.
Why it's needed
Google explains the launch by pointing to a "verification bottleneck": generative AI has greatly accelerated the creation of texts, but the number of people capable of professionally reviewing them has not increased. This year alone, the three major AI conferences (ICLR, NeurIPS, and ICML) received more than 74,000 papers. In 2025, there were about 45,000, and in 2023 — 23,000. Deep Review checks logical consistency and mathematical claims, verifies facts, and compares the work against existing research. The final conclusions remain with humans.
How the system works
Instead of a single query to a language model, Deep Review uses a four-stage agentic pipeline:
- Segmentation. A separate agent breaks the document into logical parts, such as introduction, methodology, and experiments.
- Adaptive budget. Sections with dense content, such as mathematical proofs, receive more computational resources than the introduction or literature review.
- Parallel review. Specialized agents simultaneously analyze each segment while also seeing the entire document to track cross-references.
- Synthesis and grounding. A final agent consolidates the reports, checks consistency between sections, and cross-references external sources to reduce the risk of hallucinations.
Access, API, and privacy
The service is already available worldwide. Every user gets two free document reviews per week. For larger volumes, you can connect your own Gemini API key and integrate Deep Review into workflows via the API, including CI pipelines. For now, the system accepts PDFs and plain text, and support for multiple files at once is promised for later. Google emphasizes that uploaded documents are not used for training or fine-tuning models. They are stored for one year solely to display them to the user in the interface.
Gemini 4 Argon and Antigravity
Deep Review already supports the new Gemini 4 Argon model, and according to the company, this noticeably improved analysis quality compared to earlier versions of PAT. In addition, a Teamwork Deep Document Review Pattern template has appeared in the Antigravity development environment. It replicates the Deep Review approach through multi-agent orchestration and is invoked with the /teamwork-preview command.
For developers and teams, this means that an automatic "second pair of eyes" is now available not only for code but also for technical specifications, contracts, and reports. How the tool will behave outside academic papers remains to be seen in practice, but the very idea of staged agentic review is already becoming a separate class of products.