中文导读 谷歌研究院于2026年10月14日宣布，面向公众开放人工智能工具 Deep Review：它可阅读长篇技术文档，检查数学错误、逻辑问题与缺乏依据的主张，并给出改进建议。此前该工具以 Paper Assistant Tool 之名，在 STOC、ICML、NeurIPS、ICLR 等会议试点中为逾16,000篇自愿稿件提供提交前反馈。本文说明发布背景、工作方式、使用限制与证据边界。
正文 Google Research announced on 14 October 2026 that Deep Review is now open to the public . The AI system reads long technical documents and offers careful feedback on math , logic and unsupported claims . It was first built for computer science papers . It now also checks materials in fields such as law and finance . Google says the tool can help writers improve a draft , and it can help reviewers test claims before a final decision
本段译文 谷歌研究院于2026年10月14日宣布，Deep Review 现已向公众开放。该人工智能系统可阅读长篇技术文档，并就数学、逻辑以及缺乏依据的主张提供细致反馈。它最初为计算机科学论文而设计，如今也可检查法律、金融等领域的材料。谷歌表示，该工具既可帮助作者改进草稿，也可帮助审稿人在最终决定前检验主张。
The company points to a growing problem . Generative AI helps people write papers , run code design experiments and draft proposals faster than before . Human experts , however , cannot review at the same speed According to Google , three major AI conferences received more than 74,000 submissions this year , up from about 45,000 in 2025 and 23,000 in 2023. Each paper still needs several careful reviews . That gap creates what Google calls a validation bottleneck .
本段译文 该公司指出一个日益突出的问题：生成式人工智能帮助人们比以往更快地写论文、跑代码、设计实验和起草提案，但人类专家的审阅速度跟不上。据谷歌称，今年三大人工智能会议收到逾74,000篇投稿，高于2025年约45,000篇和2023年的23,000篇。每篇论文仍需若干次认真审阅。这一差距造成谷歌所称的“验证瓶颈”。
Deep Review grew out of an earlier project called the Paper Assistant Tool , or PAT . Over the past year , Google worked with four major conferences — STOC , ICML , NeurIPS and ICLR — to give private pre-submission comments to authors of more than 16,000 manuscripts who chose to take part . Those comments were separate from official conference reviews . Papers were not used for training , and they were deleted after each program ended .
本段译文 Deep Review 由更早的项目 Paper Assistant Tool（简称 PAT）发展而来。过去一年，谷歌与 STOC、ICML、NeurIPS、ICLR 四个重要会议合作，为自愿参与的逾16,000篇稿件作者提供提交前的私下反馈。这些意见与会议正式审稿分开。论文未用于训练，并在各项目结束后删除。
The public version works in stages . First , it divides a document into logical parts . Next , it gives denser sections , such as mathematical proofs , more computing power . Then several reasoning agents check those parts in parallel while still seeing the whole text . Finally , a synthesis step compares sections and looks for outside evidence Users get two free reviews each week . People who need more can connect a Gemini API key . Uploaded files are not used to train Google models and are stored for one year only for display in the product interface .
本段译文 公开版分阶段运行：先把文档分成合乎逻辑的部分；再给数学证明等更密集的章节分配更多算力；随后多个推理智能体并行检查各部分，同时仍能看到全文；最后由综合步骤比较各节并查找外部证据。用户每周可获两次免费审阅；需要更多次数者可连接 Gemini API 密钥。上传文件不会用于训练谷歌模型，仅保存一年供产品界面展示。
Google stresses that the tool does not replace human judgment . It aims to catch logical problems , mathematical errors weak support for claims and inconsistencies between parts of a document Author feedback in the pilots was largely positive though some writers said the system sometimes marked correct proofs as wrong . For students and researchers without large review networks Deep Review may make thorough technical checking easier to reach — while final decisions remain human
本段译文 谷歌强调该工具不能取代人的判断。它旨在发现逻辑问题、数学错误、主张依据不足以及文档各部分之间的不一致。试点中作者反馈大体积极，不过也有人说系统有时会把正确证明标成错误。对缺少大型审阅网络的学生与研究者而言，Deep Review 可能让彻底的技术检查更容易获得——而最终决定仍须由人作出。
新闻来源