On October 14, 2026, Google Research released Deep Review, a public tool that uses a team of agents to check technical documents for errors. It grew out of the Paper Assistant Tool (PAT), which Google piloted at STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. Google's post says it is launching "Following successful deployments across 16,000+ submissions".
The problem it targets
Google argues that "information is being generated faster than it can be rigorously reviewed". It cites more than 74,000 submissions to ICLR, NeurIPS and ICML this year, up from about 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023. A review takes about 10 to 20 minutes, according to R&D World, and returns major and minor issues anchored to the text.
How it works
- Segmenting: an agent splits the document into sections.
- Adaptive budgeting: dense sections such as proofs get more compute than routine prose.
- Parallel verification: separate agents check each segment, each with the full document as context.
- Synthesis: a final agent consolidates findings and runs external grounding checks to cut hallucinated objections.
Deep Review now runs on Gemini 4 Argon. Google says the aim is to catch mistakes while "leaving higher-level judgements to humans". It also pitches the tool beyond papers, for design documents, presentations, and legal or financial documents. The same approach ships as a "Teamwork Deep Document Review Pattern" in Antigravity.
Access and data
Every user gets two document reviews per week free. Heavier users can connect a Gemini API key for higher limits and call a Deep Review API. Launch inputs are PDF and plain text, with multi-file input "coming soon". Google says uploaded documents are not used for training and are kept for a year only so they can be shown back to the user.
How good is it?
R&D World points to the June preprint behind PAT. On a test set of 29 known mathematical and proof errors in 26 papers, it caught 89.7%, against 55.2% for a single Gemini 3.1 Pro response. R&D World also reports that pilot users saw the system wrongly challenge correct proofs. It also reports that at Google's Research@ event, Haghtalab said the detailed checking could give researchers more time to assess "what is novel, what matters, and what kind of work actually moves the field forward."
What remains uncertain. There is no independent evaluation yet. The headline accuracy comes from a small error set, and the pilot statistics in Google's post appear only as an image we could not check as text. R&D World puts the pilot at "more than 15,000 manuscripts", slightly below Google's figure.
Analysis: The architecture is the reusable lesson: segment, budget compute by difficulty, verify in parallel with full context, then synthesise with grounding. It applies to any long document where a single model pass misses errors, from RFCs to contracts. The risk is the false objection. A reviewer that wrongly challenges a correct proof costs an author time and can mislead a tired human reviewer. Anyone adopting a tool like this should count wrong flags as carefully as caught errors.
Google Research opened Deep Review to the public on October 14 with two free reviews a week and a Gemini API option; the agentic checker segments documents, budgets compute toward proofs, verifies sections in parallel and was piloted on 16,000+ submissions at STOC, ICML, NeurIPS and ICLR.