Google Research launches Deep Review, an AI agent that checks technical documents

From conference pilot to public tool

According to Google Research, Deep Review is the evolution of its Paper Assistant Tool (PAT), an experimental project built to help researchers stress-test manuscripts before conference submission. Over the past year, Google partnered with four flagship computer science conferences, STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027, to give private, pre-submission feedback to authors of more than 16,000 manuscripts who opted in.

Google says this feedback was completely decoupled from the official conference review process. PAT did not collect any data for training, and papers were deleted from Google systems promptly after each program ended. The company describes author reactions as "extremely positive". It also says that adding Gemini 4 Argon support has raised the tool's capabilities beyond the older PAT iterations. The pilots are described in the manuscript Towards Automating Scientific Review with Google's Paper Assistant Tool.

The company frames the project as a response to a validation bottleneck: generative AI speeds up the creation of papers and technical documents, but human capacity for verification has stayed largely fixed. As an example, Google points to more than 74,000 submissions to ICLR, NeurIPS and ICML this year, up from approximately 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023.

How the four-stage pipeline works

Google argues that single-prompt LLM calls often fail on long, complex documents because they exhaust their token reasoning budget, miss cross-section dependencies, or accept nuanced claims without enough scrutiny. Deep Review instead uses a four-stage agentic pipeline:

• Document segmentation: a segmenter agent splits the document into logical segments such as introduction, experiments and methodology. Segments can overlap and must cover all of the content.
• Adaptive budgeting: the pipeline inspects information density and allocates compute per section. Sections with dense mathematical proofs may get more than an introduction or related work section.
• Deep parallel verification: specialized reasoning agents examine each segment in parallel. Each has access to the full document to track cross-references, but concentrates on its own target section.
• Global synthesis and agentic grounding: a synthesis agent merges the segment reports, runs consistency checks between sections and reruns logical checks if needed. It also performs external grounding checks and tool calls as a final safeguard against hallucinations.

According to Google, the system checks logical consistency and mathematical statements, suggests improvements, grounds facts and compares a document with existing work. It targets the most cognitively burdensome parts of review while leaving higher-level judgements to humans.

Beyond research papers

Google says the reasoning used to test whether a proof holds up can also reveal logical gaps and inconsistencies in other documents where claims need evidence. The listed use cases are:

• Design documents: catching unstated architectural assumptions, spotting edge cases and suggesting concrete improvements.
• Presentations: checking that slide transitions flow logically and that bulleted takeaways match visual charts.
• Legal and financial documents: checking internal consistency across clauses and data tables, flagging ambiguous terms or downstream consequences, and grounding claims against supporting evidence.

Google also introduced the Teamwork Deep Document Review Pattern, a review implementation inspired by Deep Review and built into Antigravity on the Teamwork multi-agent orchestration framework. It uses a multi-agent hierarchy with self-correction and segment-based rubrics to evaluate research papers, manuscript drafts and technical proposals. Antigravity users can start it with the /teamwork-preview command.

Access, limits and data policy

Deep Review is available worldwide. Every user gets two complimentary document reviews per day. Users with high-volume workflows can connect their Gemini API key for higher usage limits and programmatic integration through the Deep Review API.

PDFs and plain text are supported as inputs at launch, with multi-file input coming soon. Google states that uploaded documents are never used to train, fine-tune or improve its models. They are stored for one year solely so they can be displayed to the user in the Deep Review interface.

For authors, the release means that pre-submission critique, which Google says has generally depended on access to senior faculty, well-resourced labs or large peer networks, is open to anyone with a document. Reviewers and professionals in other fields can use it as a first check on consistency and grounding, while judgement calls remain with people.