摘要：当前颠覆人类知识获取方式的各大主流大语言模型（ LLM），从未在学术期刊上接受过独立同行评审。这一缺失值得关注。同行评审既能明晰LLM的工作原理，又有助于验证其实际效能与宣称功能是否一致。

Bring us your LLMs: why peer review is good for AI models

None of the most widely used large language models (LLMs) that are rapidly upending how humanity is acquiring knowledge has faced independent peer review in a research journal. It ’ s a notable absence. Peer-reviewed publication aids clarity about how LLMs work, and helps to assess whether they do what they purport to do.

当前颠覆人类知识获取方式的各大主流大语言模型（ LLM），从未在学术期刊上接受过独立同行评审。这一缺失值得关注。同行评审既能明晰LLM的工作原理，又有助于验证其实际效能与宣称功能是否一致。

封面报道：DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning

That changes with the publication in Nature of details regarding R1, a model produced by DeepSeek, a technology firm based in Hangzhou, China. R1 is an open-weight model, meaning that, although researchers and the public do not get all of its source code and training data, they can freely download, use, test and build on it without restriction. The value of open-weight artificial intelligence (AI) is becoming more widely recognized. In July, US President Donald Trump ’ s administration said that such models are “ essential for academic research ” . More firms are releasing their own versions.

这一局面随着深度求索公司（ DeepSeek）研发的R1模型细节在《自然》期刊的发布而改变。这家中国杭州的科技企业推出的R1是开放权重模型——虽然研究人员和公众无法获取其全部源代码和训练数据，但可以不受限制地自由下载、使用、测试并在此基础上开发。开放权重人工智能（AI）的价值正获得更广泛认可。今年7月，美国总统特朗普政府表示此类模型"对学术研究至关重要"，更多企业也开始发布自家版本。

Since R1 ’ s release in January on Hugging Face, an AI community platform, it has become the platform ’ s most-downloaded model for complex problem-solving. Now, the model has been reviewed by eight specialists to assess the originality, methodology and robustness of the work. The paper is being published alongside the reviewer reports and author responses. All of this is a welcome step towards transparency and reproducibility in an industry in which unverified claims and hype are all too often the norm.

自 1月在AI社区平台Hugging Face发布以来，R1已成为该平台解决复杂问题类模型中下载量最高的产品。如今，八位专家对该模型进行了评审，评估其创新性、方法学及稳健性。论文发表时同步公开了评审报告与作者回复。在这个充斥着未经证实宣传与炒作的行业，所有这些举措都是迈向透明度与可重现性的可喜进步。

DeepSeek ’ s paper focuses on the technique that the firm used to train R1 to ‘ reason ’ . The researchers applied an efficient and automated version of a ‘ trial, error and reward ’ process called reinforcement learning. In this, the model learns reasoning strategies, such as verifying its own working out, without being influenced by human ideas about how to do so.

深度求索的论文重点阐述了训练 R1进行"推理"的技术。研究人员采用强化学习的"试错奖励"机制高效自动化版本，使模型能自主掌握推理策略（如自我验算），而不受人类思维定式影响。

In January, DeepSeek also published a preprint that outlined the researchers ’ approach and the model ’ s performance on an array of benchmarks. Such technical documents, which are often called model or system cards, can vary wildly in the information they contain.

该公司 1月还发布了预印本论文，概述研究方法及模型在多类基准测试中的表现。这类常被称为模型卡或系统卡的技术文档，其信息含量往往差异巨大。

In peer review, by contrast, rather than receive a one-way flow of information, external experts can ask questions and request more information in a collaborative process overseen and managed by an independent third party: the editor. That process improves a paper ’ s clarity, ensuring that authors justify their claims. It won ’ t always lead to major changes, but it improves trust in a study. For AI developers, this means that their work is strengthened and therefore more credible to different communities.

相比之下，同行评审并非单向信息传递，而是由独立第三方（编辑）监督管理的协作过程：外部专家可提问并索要更多信息。该流程能提升论文清晰度，确保作者论证其主张。虽不一定引发重大修改，但能增强研究可信度。对 AI开发者而言，这意味着其工作经过淬炼，更能获得各领域认可。

Peer review also provides a counterbalance to the practice of AI developers marking their own homework by choosing benchmarks that show their models in the best light. Benchmarks can be gamed to overestimate a model ’ s capabilities, for instance, by training on data that includes example questions and answers, allowing the model to learn the correct response.

同行评审还能有效制衡 AI开发者“自评作业”的做法——他们往往会选择最能凸显模型优势的基准测试。基准测试可能存在人为操纵，例如通过包含示例问答的数据进行训练，使模型直接学习正确答案，从而高估其真实能力。

In DeepSeek ’ s case, referees raised the question of whether this practice might have occurred. The firm provided details of its attempts to mitigate data contamination and included extra evaluations using benchmarks published only after the model had been released.

在深度求索的案例中，评审专家质疑该模型是否存在此类数据污染。为此，该公司详细说明了为避免数据污染采取的措施，并补充了模型发布后新推出基准测试的评估结果。

Peer review led to other important changes to the paper. One was to ensure that the authors had addressed the model ’ s safety. Safety in AI means avoiding unintended harmful consequences, from mitigating inbuilt biases in outputs to adding guardrails that prevent AIs from enabling cyberattacks. Some see open models as less secure than proprietary models, because, once downloaded by users, they are outside of the developers ’ control (that said, open models also allow a wider community to understand and fix flaws).

同行评审还推动论文作出了其他重要修改。其一是要求作者完善模型安全性的论证。 AI安全性指避免意外危害，包括消除输出中的内在偏见、增设防护栏防止技术被用于网络攻击等。有人认为开放模型不如专有模型安全，因为用户下载后开发者即失去控制权（但值得注意的是，开放模型也让更广泛群体能够理解和修复缺陷）。

Reviewers of R1 pointed out a lack of information about safety tests: for example, there were no estimates of how easy it would be to build on R1 to create an unsafe model. In response, DeepSeek ’ s researchers added important details to the paper, including a section outlining how they evaluated the model ’ s safety and compared it with rival models.

R1的评审者指出其安全测试信息不足：例如未评估基于R1构建危险模型的难易程度。作为回应，深度求索的研究人员在论文中增补了重要细节，新增章节详细说明模型安全评估方法及与竞争模型的对比结果。

Firms are starting to recognize the value of external scrutiny. Last month, OpenAI and Anthropic, both based in San Francisco, California, tested each other ’ s models using their own internal evaluation processes. Both found issues that had been missed by their developers. In July, Paris-based Mistral AI released results of an environmental assessment of its model, in collaboration with external consultants. Mistral hopes that this will improve transparency of reporting standards across the industry.

企业正逐渐意识到外部审视的价值。上月，总部位于加州旧金山的 OpenAI和Anthropic使用各自内部评估流程互测模型，双双发现了对方开发者遗漏的问题。七月，巴黎的Mistral AI与外部顾问合作发布了其模型的环境评估报告，希望借此推动行业提升报告标准的透明度。

Given the rapid pace at which AI is developing and being unleashed on society, these efforts are important steps. But most lack the independence of peer-reviewed research, which, despite its limitations, represents a gold standard for validation.

鉴于人工智能正以惊人速度发展并深入社会生活，这些努力都是重要的进步。但大多数举措仍缺乏同行评审研究特有的独立性 ——尽管存在局限性，这种评审仍是验证体系的黄金标准。

Some companies worry that publishing could give away intellectual property — a risk, given the huge financial investment that such models have received. But, as shown with Nature ’ s publication of Google ’ s medical LLM Med-PaLM, peer review is possible for proprietary models.

部分企业担心公开发表会泄露知识产权（考虑到此类模型获得的巨额投资，这种风险确实存在）。但正如《自然》期刊发布谷歌医疗大模型 Med-PaLM所证明的，专有模型同样可以接受同行评审。

Peer reviews relying on independent researchers is a way to dial back hype in the AI industry. Claims that cannot be verified are a real risk for society, given how ubiquitous this technology has become. We hope, for this reason, that more AI firms will submit their models to the scrutiny of publication. Review doesn ’ t mean giving outsiders access to company secrets. But it does mean being prepared to back up statements with evidence and ensuring that claims are validated and clarified.

依托独立研究者进行的同行评审有助于抑制 AI行业的过度炒作。在这项技术已无处不在的今天，无法验证的宣称将对社会构成实质风险。因此我们期待更多AI企业将模型提交期刊评审：这并非要求公开商业机密，而是意味着必须用证据支撑主张，确保所有声明都经过独立验证与澄清。

来源：趣闻捕手一点号