Google Research publicly released Deep Review on October 14, 2026, making its document review system available worldwide after pilots covering more than 16,000 opted-in submissions at STOC, ICML, NeurIPS and ICLR. The system began as the Paper Assistant Tool (PAT) for computer science papers and now checks documents in fields including legal and finance, as well as Google Docs and presentations. Google says people can use it to improve drafts and verify claims during review.

The release includes two complimentary document reviews per user each week. Users with higher-volume needs can connect a Gemini API key for higher limits and programmatic access through the Deep Review API. The initial input formats are PDFs and plain text; Google says multi-file input is coming soon. Uploaded documents are stored for one year so users can view them in the Deep Review interface, and Google says the documents are not used to train, fine-tune or improve its models.

Google reported more than 74,000 submissions this year to ICLR, NeurIPS and ICML, compared with approximately 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023. Each submission requires multiple reviews, adding work for human experts. Deep Review checks logical consistency and mathematical statements, suggests improvements, grounds facts and compares a document with existing work; Google says it handles cognitively demanding review tasks while leaving higher-level judgments to people.

Before the public release, the system was tested in four conference pilots under the name PAT: STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. Authors could opt in to receive feedback in the weeks before submission deadlines. Google says the feedback was separate from official conference reviews, PAT data was not collected for training, and papers were deleted from Google systems promptly after each program ended. The company also says author feedback was positive and that the tool improved through changes to its models and PAT harness. Deep Review includes Gemini 4 Argon support, which Google says improves its capabilities beyond earlier PAT versions.

Deep Review uses four stages to review long documents. A segmenter agent divides a document into logical, potentially overlapping sections that cover its content. The system then estimates each section’s information density and allocates computing resources, giving dense mathematical proofs more budget than an introduction or related-work section. Specialized agents review different sections in parallel while retaining access to the full document for cross-references. A synthesis agent combines their reports, checks consistency across sections, repeats logical checks when needed and performs external grounding checks using tools to reduce hallucinations.

Google says the system can also review design documents for unstated assumptions and edge cases, presentations for logical flow and whether takeaways match charts, and legal or financial documents for inconsistencies, ambiguous terms and claims lacking supporting evidence. Google is also introducing a Teamwork Deep Document Review Pattern in Antigravity, built with its Teamwork multi-agent orchestration framework. It evaluates research papers, drafts and proposals using a multi-agent hierarchy, self-correction and section-based rubrics; Antigravity users can invoke it with the /teamwork-preview command. Google says support for reviewing multiple related documents, spreadsheets and exhibits together is planned.

Google Research publicly released Deep Review on October 14, 2026, making its document review system available worldwide after pilots covering more than 16,000 opted-in submissions at STOC, ICML, NeurIPS and ICLR. The system began as the Paper Assistant Tool (PAT) for computer science papers and now checks documents in fields including legal and finance, as well as Google Docs and presentations. Google says people can use it to improve drafts and verify claims during review.

The release includes two complimentary document reviews per user each week. Users with higher-volume needs can connect a Gemini API key for higher limits and programmatic access through the Deep Review API. The initial input formats are PDFs and plain text; Google says multi-file input is coming soon. Uploaded documents are stored for one year so users can view them in the Deep Review interface, and Google says the documents are not used to train, fine-tune or improve its models.

Google reported more than 74,000 submissions this year to ICLR, NeurIPS and ICML, compared with approximately 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023. Each submission requires multiple reviews, adding work for human experts. Deep Review checks logical consistency and mathematical statements, suggests improvements, grounds facts and compares a document with existing work; Google says it handles cognitively demanding review tasks while leaving higher-level judgments to people.

Before the public release, the system was tested in four conference pilots under the name PAT: STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. Authors could opt in to receive feedback in the weeks before submission deadlines. Google says the feedback was separate from official conference reviews, PAT data was not collected for training, and papers were deleted from Google systems promptly after each program ended. The company also says author feedback was positive and that the tool improved through changes to its models and PAT harness. Deep Review includes Gemini 4 Argon support, which Google says improves its capabilities beyond earlier PAT versions.

Deep Review uses four stages to review long documents. A segmenter agent divides a document into logical, potentially overlapping sections that cover its content. The system then estimates each section’s information density and allocates computing resources, giving dense mathematical proofs more budget than an introduction or related-work section. Specialized agents review different sections in parallel while retaining access to the full document for cross-references. A synthesis agent combines their reports, checks consistency across sections, repeats logical checks when needed and performs external grounding checks using tools to reduce hallucinations.

Google says the system can also review design documents for unstated assumptions and edge cases, presentations for logical flow and whether takeaways match charts, and legal or financial documents for inconsistencies, ambiguous terms and claims lacking supporting evidence. Google is also introducing a Teamwork Deep Document Review Pattern in Antigravity, built with its Teamwork multi-agent orchestration framework. It evaluates research papers, drafts and proposals using a multi-agent hierarchy, self-correction and section-based rubrics; Antigravity users can invoke it with the /teamwork-preview command. Google says support for reviewing multiple related documents, spreadsheets and exhibits together is planned.