More than 74,000 papers were submitted to ICLR, NeurIPS, and ICML this year alone. That’s up from 23,000 in 2023. Human reviewers haven’t scaled to match. Google thinks it has a partial fix, and after a year of quiet pilots, it’s making the tool available to everyone. Google announced the public release of Deep Review, an agentic document verification system that began life as an internal experiment for academic conference submissions and has since expanded into legal, financial, and general technical documents.
What Deep Review actually does
The system is built around a four-stage pipeline designed to handle the specific failure modes that trip up standard LLM calls on long, complex documents. A single prompt against a 40-page manuscript tends to miss cross-section inconsistencies, burn through reasoning budget, or let questionable claims slide. Deep Review breaks that process into structured stages: a segmenter agent first decomposes the document into logical sections, then an adaptive budgeting step allocates compute based on information density. Dense mathematical proofs get more scrutiny than a related-work section. Parallel reasoning agents then review each segment with access to the full document for context, before a synthesis agent consolidates the findings, runs consistency checks, and performs external grounding to reduce hallucinations.
The result is a tool that checks logical consistency, verifies mathematical claims, flags gaps in evidence, and compares assertions against existing work. It doesn’t replace expert judgment on high-level questions, but it handles a lot of the cognitive grunt work that makes peer review exhausting at scale.
A real track record, not a demo
This isn’t a product launching on benchmarks alone. Under its previous name, the Paper Assistant Tool (PAT), the system processed opt-in submissions across four major CS conferences: STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027. That’s over 16,000 manuscripts. Author feedback was described as strongly positive, and importantly, none of that data was used for training. Papers were deleted from Google systems after each program ended. The public version now runs on Gemini 4 Argon, which puts it ahead of the model versions used in those pilots.
Access, pricing, and format support
The public release comes with a few tiers worth knowing about:
- Free tier: two document reviews per week, no API key required
- Higher usage: connect a Gemini API key for increased limits and programmatic access via the Deep Review API
- Supported formats at launch: PDF and plain text, with multi-file input coming later
- Data policy: uploaded documents are never used for model training or fine-tuning, and are stored for one year only for display purposes within the UI
Why this matters beyond academia
The peer review framing is compelling, but the broader application is what makes this interesting for developers and enterprise buyers. The same reasoning that checks whether a mathematical proof holds also catches logical gaps in a contract, inconsistencies between a slide deck and its underlying data, or unstated assumptions in a system design document. Google is positioning Deep Review as a general-purpose verification layer for any knowledge-dense document.
Competitors in this space include tools like Elicit and Consensus on the research side, and general-purpose document analysis from Anthropic’s Claude and OpenAI’s models via custom prompting. But none of them ship a purpose-built agentic pipeline with a documented conference-scale track record. That’s a real differentiator, at least for now. Whether the quality holds up outside academic CS is the next test.
