What it is
Google Research published Deep Review on 14 October, an agentic system that reads a technical document and reports back on its reasoning, its mathematics and its supporting evidence. It grew out of the Paper Assistant Tool, an internal experiment that let authors stress-test manuscripts before a conference deadline.
The pipeline has four stages. An agent splits the document into logical sections covering all of it; a budgeting step gives more compute to dense material such as proofs than to an introduction; specialised agents verify each section in parallel, each with access to the whole document for cross-references; and a synthesis agent consolidates the findings, checks consistency between sections and runs external grounding checks. Google says higher-level judgements are left to people.
Where it has already run
As the Paper Assistant Tool, it ran with STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027. Participation was opt-in, the feedback was private and pre-submission, and it was separate from the conferences’ own peer review. Google’s research blog puts the total at more than 16,000 manuscripts; the company’s Research@ page says more than 15,000, so Google’s own materials disagree by about a thousand.
A June preprint by Google researchers reported that on a selected benchmark subset — 29 known mathematical and proof errors across 26 papers — the Paper Assistant Tool detected 89.7% of them, against 55.2% for a single Gemini 3.1 Pro response. That is a vendor evaluation of a vendor tool on a subset the vendor chose.
Why Google says it built it
The company reports more than 74,000 submissions to ICLR, NeurIPS and ICML in 2026, up from roughly 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023. Nika Haghtalab, a NeurIPS program chair, said AI has made it easier to produce papers “that look very polished, but unfortunately have weak substance”, and argued that pre-submission feedback lets authors fix weaknesses without spending reviewers’ time. At Google’s Research@ event she said detailed verification could free researchers to assess “what is novel, what matters, and what kind of work actually moves the field forward”.
Terms
Deep Review is available worldwide. The free tier is two document reviews a week; heavier users connect a Gemini API key for higher limits and programmatic access through a Deep Review API. It takes PDFs and plain text, with multi-file support listed as coming soon, and it has been extended beyond research papers to domains such as legal and finance. Google says uploaded documents are not used to train or improve its models and are stored for one year solely to display back to the user.
What to watch
The pilots measured whether authors found the feedback useful before submitting, not whether automated checking changes what conferences accept. The number worth waiting for is an evaluation run by a conference rather than by Google.
