The Pulse
Google Opens Deep Review to Technical Documents
Google Research has released Deep Review worldwide for reviewing technical and knowledge documents, including scientific manuscripts, legal agreements and financial materials. Each user receives two complimentary document reviews per week,
AI.info Team · Written with AI ·
“By providing detailed, objective evaluations of documents, Deep Review checks logical consistency and mathematical statements, suggests improvements, grounds facts, and compares with existing work.”
Rajesh Jayaram and Vincent Cohen-Addad, research scientists, Google Research
From conference papers to contracts
Google Research released Deep Review publicly on October 14, 2026, extending a tool first built to critique computer science papers to a broader range of technical and knowledge documents. The company says users can apply it to design documents, presentations, legal agreements and financial materials, as well as scientific manuscripts. Google presents the tool as an aid for drafting and verification, not a replacement for human review.
Deep Review grew out of Google’s Paper Assistant Tool, or PAT, an experimental service that offered authors private, optional feedback before conference deadlines. Google says the pilot work covered more than 16,000 manuscripts submitted to STOC, ICML, NeurIPS and ICLR. It says PAT’s feedback was kept separate from official conference review and that papers were deleted from Google systems after each pilot ended.
Two free reviews, with a one-year retention policy
The public service is available worldwide and includes two complimentary document reviews per week. Users with higher-volume needs can connect a Gemini API key for increased limits and programmatic access through the Deep Review API. At launch, the system accepts PDF and plain-text files; Google says multi-file review is planned.
Google says uploaded documents are not used to train, fine-tune or improve its models. It also says uploads are stored for one year so users can view them in the Deep Review interface. That retention period differs from the PAT pilots, where Google says it deleted papers after each program concluded.
Four stages divide the review
Google’s system breaks long documents into logical sections, then assigns review resources according to how information-dense each section is. A section containing a mathematical proof may receive more compute than an introduction or related-work section. Specialized reasoning agents examine sections in parallel while retaining access to the full document for cross-references.
A synthesis agent then combines the section-level reports, checks consistency across the document and can rerun logical checks. Google says the final stage also uses external grounding checks and tools to reduce hallucinations. The company describes the system as addressing difficulties that can arise when a single language-model prompt has to assess a long document and track dependencies across sections.
Conference submissions show the review burden
Google points to rising submission totals at major AI conferences as evidence of pressure on human reviewers. Its post says ICLR, NeurIPS and ICML received more than 74,000 submissions in 2026, compared with about 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023. The figures cover those three conferences, not all academic publishing.
The announcement does not give an independently measured accuracy rate or a comparison against human reviewers. Google says Deep Review checks claims, mathematical statements and logical consistency, but leaves higher-level judgments to people. Its public release therefore offers an additional pass over a document, while the final assessment remains with the reader.
Antigravity gets a related review pattern
Google also introduced a Deep Document Review pattern for Antigravity, its development environment. Built on the Teamwork multi-agent orchestration framework, the pattern uses a multi-agent hierarchy with self-correction and segment-based rubrics to assess papers, drafts and technical proposals. Antigravity users can invoke it with the /teamwork-preview command.
For users of the public Deep Review service, the immediate trade-off is clear: two free checks each week and support for PDFs or plain text, alongside a one-year storage period for uploaded files. Google says those files are not used to improve its models, and the company positions the resulting feedback as support for—not a substitute for—human judgment.