Google has opened Deep Review to the public after using its predecessor as an opt-in pre-submission service across more than 16,000 manuscripts for STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. Every user receives two complimentary document reviews per day; higher-volume users can connect a Gemini API key and use an API. The first release accepts PDFs and plain text, with multi-file review planned.[1]
The useful idea is not simply “ask a stronger model to review a paper”. Deep Review turns verification into a resource-allocation problem. It breaks a long document into semantic sections, estimates which parts deserve more reasoning effort, sends parallel agents into those sections with access to the whole document, then synthesizes and grounds the findings. A routine introduction and a dense proof should not receive identical computational attention.[1][2]
Four stages turn one document into a review programme
The Deep Review pipeline
| Stage | Job | Why it matters |
|---|---|---|
| 1. Segmentation | Separate theory, methods, experiments, conclusions and other logical units. | A single context window no longer has to treat every page as one undifferentiated prompt.Source 2 |
| 2. Adaptive budget | Assign more inference effort to difficult or consequential sections. | Compute is spent where a missed flaw would matter most.Source 2 |
| 3. Parallel verification | Run specialized deep-review agents while retaining full-document context. | Several claims can be checked at once without losing cross-section dependencies.Source 2 |
| 4. Synthesis and grounding | Merge, rank and de-duplicate findings and check external claims. | The final report should be coherent rather than a pile of repeated agent comments.Source 1Source 2 |
This design is an example of inference scaling: improve the result by organizing more test-time computation, not only by training a larger base model. The public release now uses Gemini 4 Argon underneath, but the published benchmark was conducted with the earlier Paper Assistant Tool architecture and Gemini 3.1 Pro. Product and benchmark should therefore be connected historically, not treated as the same frozen system.[1][2]
The 89.7% result is strong—and narrowly defined
Known equation and proof errors in the SPOT math/CS subset
| Method | Errors detected | Interpretation |
|---|---|---|
| Original SPOT state of the art | 21.1% | Earlier benchmark result under its own protocol.Source 2 |
| Gemini 3.1 Pro, zero shot | 55.2% | One direct model-based review under the study’s evaluation.Source 2 |
| Paper Assistant Tool | 89.7% | Inference-scaled pipeline using the same model family.Source 2 |
The benchmark supports a specific conclusion: an orchestrated review process found more of those known errors than a single zero-shot call. It does not establish 89.7% accuracy for every legal contract, financial report, experimental paper or presentation. Recall also does not describe how many valid passages were incorrectly flagged. Authors still have to investigate the critique rather than accept it as a verdict.[2]
Public access changes the review bottleneck—but does not remove it
The pilots show scale. Google says the system has processed more than 16,000 conference submissions, far beyond the 4,700-plus STOC and ICML manuscripts described in the original paper. That creates a new layer before formal review: authors can test claims, equations and experimental descriptions while revisions are still possible. It can save reviewer attention when it catches a real defect; a false alarm consumes expert time instead.[1][2]
The privacy policy is unusually concrete. Google says uploaded documents are not used to train, fine-tune or improve its models, and are stored for one year solely so they can be displayed in the Deep Review interface. That retention period matters for unpublished research, contracts and internal financial material; users still need to decide whether their own confidentiality obligations permit the upload.[1]
The deeper lesson is about verification design
Most people use an AI reviewer as a single conversation: paste the document, ask for flaws, read the answer. Deep Review treats that workflow as underpowered. It allocates attention, creates parallel lines of attack, preserves document-wide context and then asks a separate process to reconcile the findings. The architecture can be copied conceptually even when the product is not used: divide a high-stakes review into explicit checks, give difficult claims more effort, and require evidence for every serious objection.
The public launch makes technical review more accessible, but its success will depend on a human discipline the system cannot automate away: distinguishing a persuasive critique from a correct one. Deep Review’s real contribution is not an artificial referee. It is a structured way to spend machine attention before human judgment becomes expensive.
Sources & further reading
- Analyzing technical documents with Deep Review Google Research · 14 Oct 2026 · Accessed 16 Oct 2026
- Towards Automating Scientific Review with Google’s Paper Assistant Tool arXiv · 26 Jun 2026 · Accessed 16 Oct 2026