Google Research launches Deep Review publicly: a four-stage agentic document verifier, piloted on 16,000+ conference papers, now on Gemini 4 Argon with an API

Google Research made Deep Review public on 2026-10-14. It evolved from the Paper Assistant Tool (PAT), which gave opt-in, pre-submission feedback on more than 16,000 manuscripts at STOC 2026, ICML 2026, NeurIPS 2026 and ICLR 2027, fully decoupled from official review. The motivation is a verification bottleneck: Google says ICLR, NeurIPS and ICML received over 74,000 submissions this year, up from about 23,000 in 2023.

Instead of a single long prompt, Deep Review runs a four-stage agentic pipeline: a segmenter agent splits the document into overlapping logical segments; adaptive budgeting gives denser sections such as proofs more compute; specialized agents verify segments in parallel while seeing the full document for cross-references; and a synthesis agent consolidates findings, reruns cross-section checks and uses external grounding tools against hallucination. It now supports Gemini 4 Argon. No head-to-head benchmark scores were published; the PAT methodology and pilot statistics are in arXiv 2606.28277.

Every user gets two free reviews per week; connecting a Gemini API key raises limits and enables the Deep Review API for CI and drafting workflows. PDF and plain text are supported at launch, with multi-file input coming. Uploads are never used to train Google models and are kept for one year for display. Antigravity also gains a Teamwork Deep Document Review pattern via /teamwork-preview.