Google Deep Review: Our Expert Review of Google’s AI Paper Checker

5/5 - (2 votes)

Google Deep Review is a free AI reviewer from Google Research. You give it a long technical document, and it hunts for errors in the math, the logic, the stated facts and the experiments. It grew out of the Paper Assistant Tool (PAT), which checked more than 15,000 conference papers before Deep Review opened to everyone on October 14, 2026.

Our verdict is 8 out of 10. Deep Review is the best free pre-submission check for researchers and technical writers today. It catches proof bugs that a normal chatbot misses. Still, it makes false calls, reads one PDF or text file at a time, and gives you only two free reviews a week.

What Google Deep Review is

Deep Review is an agentic AI system for feedback and technical checks on complex documents. Google Research built it first for computer science papers. It now also covers design documents, slide decks, and legal and financial files.

The tool exists because of a simple problem. AI now helps people write papers faster than humans can check them. ICLR, NeurIPS and ICML got over 74,000 submissions this year. In 2023 the same three had about 23,000.

Deep Review does not score a paper or predict acceptance. It reports errors and suggests fixes, and it leaves the big judgment calls to people. The PAT research paper puts it plainly: the system does not produce subjective assessments or rankings.

So who is it for? It is for anyone who must be right before they publish:

  • Paper authors: Researchers can catch a broken lemma or a weak experiment before a referee does. This is the job the tool was built for.
  • Engineers: Teams can run a design document through it to find unstated assumptions and edge cases. That matters most when a spec is long and many people rely on it.
  • Legal and finance teams: Deep Review checks that clauses and data tables agree with each other. It also flags vague terms and statements that lack evidence.
Google Deep Review upload page for checking technical documents
Screenshot: research.google

How Deep Review works

A single prompt to a large model often fails on a long document. The model runs out of reasoning budget, misses links between sections, or accepts a tricky statement without a fight. Deep Review splits the work across many agents to avoid that.

The four-stage Deep Review pipeline

  1. Document segmentation: A segmenter agent cuts the document into logical parts, such as the method, the proofs and the experiments. Parts may overlap, and together they must cover every page.
  2. Adaptive budgeting: The pipeline measures how dense each part is. A page of proofs gets far more compute than the related work section, through light, medium or high thinking tracks.
  3. Parallel verification: Specialist reasoning agents check each part at the same time. Each agent sees the whole document, so it can follow cross-references, but it focuses on its own section.
  4. Global synthesis: A final agent merges the reports, removes duplicates and rates severity. It then runs consistency checks across sections and uses search to catch made-up references.

Gemini models do the reasoning. The pilots ran on an advanced version of Gemini 2.5 Deep Think, and the public tool now supports Gemini 4 Argon.

The input is a PDF or plain text. Multi-file input is coming soon, but it is not live yet. The output is a written review of the document’s logic, math, facts and experiments, with suggested fixes and a comparison to existing work.

Deep Review four-stage agent pipeline diagram on the Google Research blog
Screenshot: research.google

Deep Review features that matter in practice

Under the pipeline, a handful of features decide how useful a review is day to day:

  • Proof and math checks: Deep Review tests mathematical statements step by step. At STOC it flagged wrong inequality use, calculation slips and gaps in proofs, and one author called a find “a critical bug” that broke the whole proof.
  • Experiment review: It questions benchmark results, asks for missing ablations and spots mismatched setup details. At ICML, 31% of surveyed authors ran new experiments after reading the feedback.
  • Reference grounding: The synthesis agent searches for cited work to catch fake or wrong references. This is a real gain over plain chat models, which can invent citations.
  • Prior work comparison: Deep Review compares your results with existing papers. That helps you spot a “new” result that someone already published.
  • Custom context: You can steer a review toward a target, such as the math or a specific conference track. A focused review wastes less effort on parts you already trust.
  • Review depth and thinking level: Review depth sets how many multi-agent review and verification passes run (Low, Medium, High or Extra High). Thinking level sets how much reasoning each Gemini call spends (Low, Medium or High). Free reviews run on Medium for both, and your own API key unlocks every combination.
  • Gemini API key and Deep Review API: Heavy users can connect a Gemini API key for reviews beyond the free two a week. Deep Review also has its own API (deepreview v1) in the Google APIs Explorer, so labs can wire checks into their writing workflow.
  • Antigravity pattern: Google built a Teamwork Deep Document Review Pattern into Antigravity, run with the /teamwork-preview command. Developers get the same multi-agent review idea inside their coding tool.
  • No training on uploads: Uploads are not used for model training without your permission. Google stores them for up to one year so you can see past reviews, then deletes them.

Deep Review benchmark results and pilot data

Google tested PAT, the engine behind Deep Review, on the SPOT benchmark. The test set was a math and computer science subset with 26 papers and 29 known errors in equations and proofs.

MethodError detectionErrors found (of 29)
Original SPOT best result21.1%about 6
Gemini 3.1 Pro, single prompt55.2%16
PAT on Gemini 3.1 Pro89.7%26

The jump from 16 to 26 caught errors is the core case for the agent design. The same model found 10 more errors once the work was split and checked in depth. Keep two caveats in mind, though. The test set is small, and the team used a modified grader, so the numbers do not line up exactly with the original SPOT results.

Real users tell a fuller story. Authors at STOC 2026 and ICML 2026 answered a survey after their reviews. Across both conferences, PAT checked more than 4,700 submissions.

Survey resultSTOC (124 replies)ICML (733 replies)
Would use PAT again97%92.1%
Very or mostly helpful92.7%90.7%
Paper became easier to read85.1%87.0%
Has teaching value for students75.2%83.9%
Feedback mostly or all grounded55.8%64.8%
Found real gaps in theory11.6%35.4%
Ran new experimentsNot asked31%

The helpful scores are high. The grounding row is the one to watch. In both groups, a third or more of authors felt that a good part of the feedback was not fully grounded. In other words, Deep Review is a sharp second reader that still needs a human to check every flag.

Paper Assistant Tool research paper abstract on arXiv
Screenshot: arxiv.org

The pilots later grew to NeurIPS 2026 and ICLR 2027. At ICLR 2027, authors could opt in through OpenReview, and reviewers never saw the feedback.

How to use Google Deep Review step by step

Running a review takes a few minutes of setup and some waiting:

  1. Prepare one file: Export your document as a single PDF, or copy the plain text. If your work has an appendix in a separate file, merge it first, because multi-file input is not live yet.
  2. Keep the PDF light: Google publishes no size cap for Deep Review. The ICLR program asked for PDFs under 20MB, ideally under 10MB, so that is a safe target.
  3. Open the tool: Go to the Deep Review page and upload the PDF or paste the text.
  4. Add custom context: Tell it what to focus on, such as the proofs in section 3 or a target venue. Short, specific notes work best.
  5. Submit and wait: Results appear in the Your documents table when the run finishes. With an API key, Deep Review shows an estimated cost first, so you can decide before it starts. Your two free reviews reset each week.
  6. Triage the report: Read the most serious issues first, then the smaller fixes. Mark each item as confirmed, false alarm or needs a second look.
  7. Scale with an API key: If you need more than two reviews a week, connect a Gemini API key. You then pay only the standard Gemini API token rates for each extra review, and Deep Review shows an estimated cost before the run starts. There is no extra charge for the tool itself.

Deep Review use cases for researchers, SEO and content teams

Researchers and paper authors

This is home ground. Run Deep Review a week or more before a deadline, so you have time to fix a proof or rerun an experiment. Use it again on the final draft to catch new typos and broken references. It is also useful for checking a draft from a student before you send it on.

SEO and content teams

Most SEO work does not need this tool. Keyword research, rank tracking and short blog posts are outside its scope. But many content teams now publish original research to earn links and AI citations, and that is where Deep Review helps.

Take a 30-page industry study with charts, survey data and a method section. Deep Review can check that the numbers in the text match the tables, that the method supports the headline findings, and that cited studies exist. One wrong percentage in a widely linked study can follow a brand for years, so a free check before launch is cheap insurance.

Whitepapers, technical documentation and long B2B guides are other good candidates. Expect weaker results on chart-heavy pages, because figures and complex notation were a weak spot in the STOC pilot.

Publishers and legal or finance teams

Editors of technical journals and trade publications can use it to screen long submissions for internal contradictions. Legal and finance teams can check that a contract’s clauses agree, or that a report’s tables match its text. In each case, a human expert still makes the final call.

Where Deep Review does not fit

  • Open web research: Deep Review audits a document you already have. It does not go out and build a report on a topic for you.
  • Writing first drafts: It suggests fixes, but it does not write or rewrite your content.
  • High-volume agency work: Two free reviews a week run out fast across many clients. Without a published API price, budgeting for more is guesswork.
  • Formal peer review: The pilots kept PAT outside the official review process on purpose. Using it as a judge of someone else’s paper goes beyond what it was tested for.

Deep Review limitations and risks

Every flag in a Deep Review report is a lead to check, and these are the main ways it can mislead you:

  • False alarms: The model sometimes marks a correct proof or argument as wrong. The PAT team calls this a problem of any LLM-based system, so a confident flag can still be a false alarm.
  • Date and knowledge errors: Early pilots saw wrong dates and outdated knowledge. Better search tools reduced this, but recent results in a fast field may still trip it up.
  • Parsing trouble: PDF parsing caused errors in the pilots, and complex notation and figures confused the model at STOC. Clean, text-based PDFs give better results than scanned pages.
  • Thin public evidence: The benchmark covers just 29 errors, and Google has published no accuracy numbers for legal, finance or design documents. Treat use outside research papers as early.
  • One-year storage: Uploads stay on Google systems for a year. That may clash with client contracts or embargoed research, so check your own rules before you upload.

Deep Review vs Gemini Deep Research, ChatGPT deep research and Gemini Notebook

Deep Review is easy to confuse with “deep research” tools. They do opposite jobs. Research tools gather sources and write a report for you, while Deep Review checks a document you wrote.

ToolMain jobWhat you give itWhat you getAccess
Google Deep ReviewFind errors in a finished documentOne PDF or plain textError report with fixes, grouped by severityTwo free reviews a week worldwide; API key for more
Gemini Deep ResearchResearch a topic and write a reportA prompt, plus optional files, Gmail, Drive or ChatMulti-page report, also as an Audio Overview150 countries, 45+ languages; free Pro plan for college students for one year
ChatGPT deep researchResearch a question across many sourcesA prompt, uploaded files and connected appsCited report you can download as Markdown, Word or PDFDepends on ChatGPT plan and country
Gemini Notebook (formerly NotebookLM)Work with and analyze your own sourcesYour source material in a notebookAnswers and analysis grounded in your sources, plus code-based data analysis on some plansStandalone or in the Gemini app; code tools on AI Ultra first

The smart setup uses two tools. Gather and draft with a research tool, then run the finished document through Deep Review before it goes out. Gemini Notebook helps you think inside your sources, while Deep Review tries to break your argument.

Deep Review pros and cons

Pros

  • Free real depth: Two full multi-agent reviews a week cost nothing. You need no API key and no paid plan to start.
  • Proven on real papers: The design was tested on thousands of live submissions at STOC, ICML, NeurIPS and ICLR. That is a large field test for a free tool.
  • Focused on facts: It skips accept or reject guesses and sticks to errors and fixes. That makes the output easier to act on.
  • Clean privacy terms: No training on uploads is a firm promise. That removes the biggest worry for unpublished work.

Cons

  • Single-file input: Projects split across a main paper, appendix and code have to be merged first. That is extra work for large submissions.
  • Unknown cost at scale: Google has published no price for API reviews. Teams cannot plan a budget for heavy use.
  • Narrow audience: Marketing copy, short posts and casual content gain little from it. The tool pays off only on long, technical, high-stakes documents.

Google Deep Review verdict and rating

Rating: 8/10. Deep Review is the most useful free tool for checking a technical document before it goes public. The agent design caught 26 of 29 known errors where a single prompt caught 16, and over 90% of pilot authors found it helpful. Those are rare numbers for a free tool.

It loses points for false alarms, partly ungrounded feedback, one-file input and a hidden price above the free tier. It is a second reader with a sharp eye, and it does not replace a human expert.

Use it if you write papers, technical reports, original research studies or long specs. Skip it if you need web research, writing help or volume checks across many short pieces. For SEO and content teams, save your two weekly reviews for the studies and whitepapers that carry your brand’s name.

Add Comment