Towards Automating Scientific Review with Google's Paper Assistant Tool

Google researchers have introduced the Paper Assistant Tool (PAT), an agentic AI framework designed to automate deep scientific review and verification of full research manuscripts, according to a paper submitted on 26 June 2026 [1]. The system ingests complete scientific papers and produces evaluations that check theoretical results, validate experiments, and identify potential flaws [1]. By using inference scaling techniques, PAT achieved a 34% improvement over zero-shot recall on mathematical errors in the SPOT benchmark [1]. The framework was piloted as a pre-submission tool for authors at two major computer science conferences, STOC and ICML, where it identified critical errors and suggested substantive improvements [1]. The researchers argue that traditional human peer review cannot scale to match the accelerating pace of AI-assisted science, and that AI must be deployed to accelerate verification itself [2]. The project arrives amid broader scrutiny of AI reliability. Large language models, the technology underpinning tools like PAT, are known to produce hallucinations—responses containing false or misleading information presented as fact [5]. Detecting and mitigating such errors remains a significant challenge for deployment in high-stakes scenarios [5]. Google’s own Gemini chatbot family, which shares underlying research lineage with the company’s AI efforts, faced a suspension of its image-generation capabilities in early 2024 after users reported historical inaccuracies and bias [9]. Subsequent Gemini updates through 2025 focused on reducing hallucinations and improving agentic capabilities for autonomous research [9]. The PAT paper proposes a taxonomy of four progressive levels of AI-human collaboration in scientific evaluation, framing the tool as a way to ease the cognitive burden on referees while preserving human control over review outcomes [2]. The authors position the work as a step toward resolving the tension between the volume of AI-generated research and the capacity of human reviewers [2]. The submission appears on arXiv under the machine learning category [1].