Google Opens Deep Review for Public Technical Document Checks
The agentic reviewer is now publicly available for PDFs and plain text, but its launch evidence comes from Google-run pilots and an authors' preprint.
Edited by Tyronne Panaino
Google Research publicly launched Deep Review on October 14, opening an agentic system for technical feedback and document verification to users worldwide. The launch targets researchers, engineers and other knowledge workers who need to examine complex material, with initial support for PDF and plain-text files.
The change matters because Google is moving a system previously used in opt-in conference pilots into a general service. That creates a new way to stress-test manuscripts and other technical documents, but it does not make the system an authority on whether a paper, contract or financial analysis is correct. Google's own design keeps higher-level judgment with people, and the available performance evidence comes from the system's developers.
What the public launch includes
The Google Research launch article says every user receives two complimentary document reviews per week. Higher-volume users can connect a Gemini API key for additional limits and programmatic access. At launch, the service accepts PDFs and plain text; Google says multi-file review is still coming.
The data policy deserves equal attention. Google says uploaded documents are not used to train, fine-tune or improve its models. It also says those documents are stored for one year so they can be displayed in the Deep Review interface. Teams assessing unpublished research, confidential designs or regulated documents therefore need to evaluate that retention period against their own information-handling rules before uploading material.
How the review pipeline works
Google describes a four-stage process intended to spend more computation where a document is hardest to verify. A segmenter first divides the file into logical, potentially overlapping sections. An adaptive budgeting stage then allocates different amounts of compute based on information density, so a mathematical proof can receive more attention than an introductory section.
Specialized reasoning agents examine those segments in parallel while retaining access to the full document for cross-references. A synthesis agent combines the reports, checks consistency across sections and performs external grounding steps. Google positions this architecture as an answer to limitations it sees in single model calls over long documents, where dependencies can be missed or a reasoning budget can be exhausted. The launch page says Deep Review now includes Gemini 4 Argon support.
Those mechanics explain the product's intended advantage, not a guarantee that every warning is valid or every error will be found. Users still need to inspect cited evidence, reproduce calculations and decide which suggestions are relevant to the document's purpose.
What the pilot evidence establishes
Google says the earlier Paper Assistant Tool was offered to authors who opted in around STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027, covering more than 16,000 manuscripts. The feedback was separated from the conferences' official review decisions. That history shows substantial use in a pre-submission setting, but it does not by itself measure how the public service performs across legal, financial, design or presentation documents.
The authors' June preprint documents an earlier version of the framework and reports a 34% improvement over a zero-shot baseline for recall of mathematical errors on the SPOT benchmark. It also describes pilots at two computer-science conferences. The result is narrow: it concerns one benchmark measure from the system's authors, not an independent comparison of false positives, missed defects, latency, cost or outcomes across the broader document classes now promoted at launch.
What to watch next
The useful checkpoints are independent evaluations on representative long documents, transparent false-positive and miss rates, and evidence that citations and external grounding remain reliable across domains. Multi-file support will create another test because consistency checks across spreadsheets, exhibits and related documents are more demanding than reviewing one file.
Status
Confirmed. Internal confidence is medium. Google's official pages establish the public release, access terms, data policy, architecture and pilot history, while the benchmark evidence is author-reported and this run found no independent reliability study for the launched service.
Sources
- Google Research — Analyzing technical documents with Deep Review
- arXiv — Towards Automating Scientific Review with Google's Paper Assistant Tool
Update note: Last reviewed 2026-10-15. We will revise this post if Google changes availability or retention terms, adds multi-file support, or independent evaluations test the public system.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.