Analyzing technical documents with Deep Review

Google Research announced the public release of Deep Review, an agentic AI system for rigorous feedback and technical verification on complex documents. The announcement comes from research scientists Rajesh Jayaram and Vincent Cohen-Addad, following pilot deployments across more than 16,000 submissions at STOC, ICML, NeurIPS, and ICLR. The system, previously known as the Paper Assistant Tool (PAT), is now available worldwide through the Google Research Portal.
Deep Review was initially built for computer science academic paper analysis but has expanded to review documents across domains including legal and finance. It functions both as an iterative improvement tool during drafting and as a verification instrument for checking claims and evaluation during the review process. The system was developed collaboratively across Google Research, Google DeepMind, and Google Cloud.
Access is tiered. Every user receives two complimentary document reviews per week. Users with high-volume workflows can connect their Gemini API key for higher usage limits and programmatic integration via the Deep Review API. Initially, the system supports PDFs and plain text as input formats, with multi-file inputs coming soon. On privacy, Google states that documents uploaded to Deep Review are never used to train, fine-tune, or improve its models, and uploaded documents are stored for one year for the sole purpose of display to the user in the Deep Review UI.
The system addresses what Google describes as a validation bottleneck: AI has accelerated content creation in scientific papers, code, experiments, and proposals, while human capacity for verification has remained largely fixed. Google points to submission growth at major AI conferences as evidence. This year saw over 74,000 submissions across ICLR, NeurIPS, and ICML, up from approximately 45,000 in 2025, 32,000 in 2024, and 23,000 in 2023. Each submission requires multiple high-quality reviews, placing a growing burden on human experts.
During the four pilot partnerships, authors could opt in to receive feedback and analysis on their papers in the weeks before submission deadlines. This feedback was completely decoupled from the official conference review process. PAT did not collect data for training, and papers were deleted from Google systems promptly after each program concluded. Google reports extremely positive author reactions, and says continued improvements to underlying models and the PAT harness have increased the tool's capabilities, with the addition of Gemini 4 Argon support pushing it beyond older PAT iterations.
Deep Review uses a four-stage agentic pipeline. First, a segmenter agent decomposes the document into logical, possibly overlapping segments that together cover all content, such as introduction, experiments, and methodology. Second, adaptive budgeting inspects information density and dynamically allocates a compute budget to each section, with a specialized inference scaling pipeline; dense mathematical proofs may receive more compute than an introduction. Third, deep parallel verification has specialized reasoning agents scrutinize each segment in parallel, each with access to full document context to track cross-references while focusing reasoning power on its target section. Fourth, a synthesis agent consolidates segment-level reports, performs final consistency checks between sections and reruns logical checks if needed, and conducts external grounding checks and tool calls as a safeguard against hallucinations.
The rationale for the staged design is that single-prompt LLM calls often fail on long, complex documents by exhausting their token reasoning budget, overlooking long-context cross-section dependencies, or accepting nuanced claims without full scrutiny. The pipeline checks logical consistency and mathematical statements, suggests improvements, grounds facts, and compares against existing work, while leaving higher-level judgments to humans.
Google argues the same long-context reasoning used to check whether a mathematical proof holds up can surface logical gaps in other document types. Example uses include design documents, where it can catch unstated architectural assumptions and spot edge cases; presentations, where it can check that slide transitions flow logically and bulleted takeaways match visual charts; and legal and financial documents, where it can check internal consistency across clauses and data tables, flag ambiguous terms or downstream consequences, and ground claims against supporting evidence. This suggests the target audience extends beyond academic researchers to knowledge workers in engineering, legal, and finance functions.
Google is also introducing the Teamwork Deep Document Review Pattern, a document review implementation pattern inspired by Deep Review and built into Antigravity using the Teamwork multi-agent orchestration framework. The pattern is designed to evaluate research papers, manuscript drafts, and technical proposals using a multi-agent hierarchy with self-correction and segment-based rubrics. Antigravity users can invoke it with the /teamwork-preview command.
Google frames the broader mission around access: comprehensive pre-submission critique in computer science has generally depended on review networks such as senior faculty, well-resourced laboratories, or peer advisory networks. Automated logical verification is positioned as a way to help academics stress-test assertions, identify avoidable flaws, and strengthen documents. Google says it will continue expanding the platform, citing Gemini 4 Argon support and upcoming multi-file support that will let users run consistency checks across related documents, spreadsheets, and exhibits simultaneously.
The practical constraints are worth noting. In practice, two weekly reviews is enough for occasional use but not for high-volume pipelines, which is why the API path tied to a Gemini API key exists. The one-year retention window for uploaded documents may matter to organizations with strict data handling policies, since it applies even though the content is not used for training. A likely trade-off is that as the system moves from mathematical and scientific review into legal and financial domains, grounding and hallucination safeguards become more consequential, because errors in those settings can carry direct cost. The company does not provide public benchmark numbers in this announcement, so claims about quality beyond the conference pilots rest on the reported author feedback and pilot scale.
Why it matters: Researchers and technical teams drafting papers, design docs, or proposals now have a free, API-accessible tool for automated consistency and grounding checks, which could reduce the time spent on mechanical review and shift human effort toward judgment calls. For teams building review workflows, the Deep Review API and the Antigravity /teamwork-preview pattern offer two integration paths, one programmatic and one embedded in a multi-agent orchestration framework. Inference: adoption will likely be highest where documents are long and claims are checkable, while regulated domains may weigh the one-year storage policy carefully before uploading sensitive material.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
General Science