Accessibility Adjustments

Use these optional tools to adjust reading and display preferences. These tools cannot resolve every accessibility barrier. Please contact the website owner if you need assistance.

  • Text adjustments
  • Content scaling 100%
  • Font size 100%
  • Line height 100%
  • Letter spacing 100%
  • Colour adjustments
  • Orientation adjustments

Google opens Deep Review for technical document checks

Listen to this article

Google released Deep Review publicly on October 14, opening its agentic document checking system beyond research conference pilots. The service reviews technical arguments and supporting evidence, with Gemini 4 Argon among its current capabilities.

Access and document handling

Google says access is worldwide, with two complimentary document reviews per user each week. PDFs and plain text are supported initially. Multiple file inputs are planned. Higher usage limits and programmatic integration use a Gemini API key. The announcement gives no specific paid review price.

The entry point is the Deep Review portal. Google says uploads are kept for one year to display them in the interface and are never used to train, fine tune or improve its models.

How the review is divided

The underlying Paper Assistant Tool research, published in June, describes a system that splits a manuscript into logical segments and allocates more computation to demanding sections. Review agents examine their assigned sections while retaining the full document as context. A synthesis step combines findings and uses search to check references and reduce invented claims.

What the evidence measures

In a narrow benchmark of 26 mathematics and computer science papers containing 29 equation or proof errors, the authors report that PAT with Gemini 3.1 Pro detected 89.7% of the known errors, compared with 55.2% for a single model response. The grader was adapted to recognize logically equivalent error descriptions, with author audits. These are developer reported results for the earlier system, rather than a measurement of the newly released service.

The paper also records failure modes including incorrect dates, PDF parsing trouble and critiques that mistakenly label valid arguments as wrong. Its authors say search and parsing improvements addressed the first two. Reasoning mistakes remained an active problem.

The official ICML retrospective gives useful context for the pilot feedback. About 4,500 papers received private assistance before submission. Among survey respondents whose papers contained theory, 35.4% reported substantial gaps that took over an hour to fix. Among those with experimental results, 31% said they ran new experiments. Those percentages describe respondents and relevant paper types, not all submissions. The feedback was separate from formal peer review and unavailable to conference reviewers.

A separate Antigravity route

Google also announced a Deep Document Review pattern for Antigravity, inspired by Deep Review.

The Teamwork documentation places document review among its supported execution paths. The teamwork preview command is available on paid Antigravity plans in the app and CLI. Users describe the goal, go through a scoping interview and approve the resulting brief before execution. This is a separate access route with its own plan requirements. The portal’s two free weekly reviews should not be assumed to apply to it.

ByteForward verified the public announcement and documentation but did not submit a document or test review quality.

The featured photograph shows Google headquarters in 2009. Photo by Steven Arnold under CC BY 3.0. Converted to WebP. It is archival company context.

Maya Chen
Maya Chen

Maya Chen is focused on covering AI models, research, and the evidence behind new capabilities. Maya follows model launches, benchmarks, open weights, and scientific uses of AI with one question in mind. What changed, and how would we know? The voice is curious and exacting, with a soft spot for elegant technical ideas and little patience for a leaderboard without context.

Leave a Reply

Your email address will not be published. Required fields are marked *

Gravatar profile