内容来源：https://www.sciencenews.org/article/ai-model-reason-logic-words-efficient

内容总结：

AI能否不借助语言进行推理？一个小型模型给出了新思路。最新研究表明，跳过语言环节可能让部分AI推理更高效。

目前许多AI系统在处理复杂问题时，会先输出大量中间步骤的文字链，再给出最终答案，这种“思维链”推理方式虽然能提升部分任务表现，但意味着模型在得出简短答案前需生成大量词元，每个词元都消耗算力，导致推理速度更慢、成本更高。

一个名为BDH-CQ的小型实验性AI系统采取了不同路径。研究人员在8月10日提交至arXiv的论文中报告，该系统能在不明确写出中间思考过程的情况下解决部分推理谜题。研究探讨了AI是否需要在每一步推理中都依赖语言，以及更多内部计算能否降低部分推理成本。

在训练中，AI模型通常会先看几个示例再尝试新问题，许多模型在工作时会一直参考这些示例。但在推理测试中，BDH-CQ将每个示例用于更新一个固定大小的记忆，该记忆在接收更多示例时保持不变，因此无需反复回溯已看过的内容。AI公司Pathway首席执行官、复杂性科学家祖赞娜·斯塔米罗斯卡表示：“查询到来后，模型完全不用文字写出思考过程，中间没有任何环节转化为语言。”

斯塔米罗斯卡和同事称，该模型在公开的ARC-AGI-1评估集上尝试两次时，能解出近三成谜题。这表明小型模型可以在不写出中间步骤的情况下解决部分新推理问题。该基准测试要求AI系统从少量示例中推断视觉规则并应用于新谜题，类似学生考试。

该模型在不同谜题上表现各异，例如能处理部分涉及图形旋转或移动的任务，但在颜色变化和部分规则组合上表现较差。在更难的排序和嵌套谜题上，模型在看到难度相似的示例后表现有所提升。研究人员估计，每次谜题查询运行成本约0.0007美元，约为GPT-5.6 Luna在同一基准测试上的十一分之一，但两者成本计算方式不同。该研究尚未经过同行评审。

加拿大滑铁卢大学计算机科学家邓云天称这一发现是“一个有趣的效率结果”，但表示这并不能证明BDH-CQ的底层架构优于其他方法。他说，像BDH-CQ这样的专用小型模型是实现高效推理的有前景路径，但需要更多测试来区分模型设计效果与训练方式的影响。没有写出的推理过程使模型更难被检查，不过即使有文字思维链，也未必能忠实反映模型得出答案的真实过程。

德国ELLIS研究所蒂宾根分部及马克斯·普朗克智能系统研究所的机器学习研究员约纳斯·盖平指出，BDH-CQ专为ARC类问题构建，难以与通用AI系统直接比较，但他称该方法“很巧妙”，并指出它无需重新训练即可应对测试问题。对邓云天而言，这项研究提出了一个更广泛的问题：AI是否必须将所有推理都转化为语言？他说：“人类语言有助于交流推理，但未必是每步中间计算的最有效表示方式。”

中文翻译：

人工智能能否不借助语言进行推理？一个小模型对此进行了验证

跳过语言环节或许能让某些人工智能推理更高效

人工智能或许能够在不必将每一步都诉诸语言的情况下进行推理。

如今许多人工智能系统在解决难题时，会先通过生成一长串文字形式的中间步骤，再得出答案——这种技术通常被称为思维链推理。这种方法可以提升某些任务的表现，但也意味着人工智能模型在得出相对简短的答案之前，可能会生成大量文字，即词元。每生成一个词元都需要消耗算力，因此推理过程越长，答案生成就越慢、成本也越高。

一个名为BDH-CQ的小型实验性人工智能系统采取了不同的方式。研究人员在8月10日提交至arXiv.org的一篇论文中报告称，它能在不明确写出中间思考过程的情况下解决一些推理谜题。这项工作提出了一个问题：人工智能是否在推理的每一步都需要语言，以及更多地在内部完成推理是否能让某些推理变得更低成本。

在训练过程中，人工智能模型通常先看几个示例，然后再尝试解决新问题。许多模型在工作时会一直将这些示例摆在面前。但在推理测试中，BDH-CQ（DH是“龙之幼崽”的缩写）则利用每个示例来更新一个固定大小的记忆，该记忆在接收更多示例时保持同样的大小，因此它不必反复回顾已经看过的所有内容。

“一旦查询到来，模型根本不会用文字写出它的思考过程，”人工智能公司Pathway的首席执行官、复杂性科学家祖赞娜·斯塔米罗夫斯卡说。斯塔米罗夫斯卡表示，该模型在内部完成问题求解，不会将每一步转化为文字。“中间没有任何东西会转化为语言，”她说。

斯塔米罗夫斯卡及其同事表示，在公开的ARC-AGI-1评估集上，该模型在获得两次尝试机会的情况下解出了近三成的谜题。这一结果表明，一个小模型可以在不写出中间步骤的情况下解决一些新的推理问题。该基准测试要求人工智能系统从少量示例中推断出视觉规则，并将其应用于新的谜题。这些测试类似于人类学生的考试。

该模型在某些谜题上表现优于其他谜题。例如，它能处理一些涉及旋转或移动形状的任务，但在颜色变化和某些规则组合方面则较为吃力。在较难的排序和嵌套谜题上，该模型在看过一个难度相近的示例后表现有所提升。

斯塔米罗夫斯卡及其同事表示，用词更少的过程可能有助于降低计算成本。他们估计，每次谜题查询的运行成本约为0.00070美元，大约是GPT-5.6 Luna在同一基准测试上成本的十一分之一，不过两者的成本计算方式不同。该研究尚未经过同行评审。

加拿大滑铁卢大学的计算机科学家邓云田表示，这一发现是“一个有趣的效率结果”。但他指出，这并不能证明BDH-CQ的底层架构优于其他方法。

“像BDH-CQ这样的专用小模型是通往高效推理的一条有前景的路径，”邓说。但他表示，需要更多测试来区分模型设计的效果与其训练方式的效果。邓指出，没有写出的推理过程，模型更难被审查。但即便是写出的思维链，也可能无法忠实展示模型是如何得出答案的。

德国图宾根ELLIS研究所和马克斯·普朗克智能系统研究所的机器学习研究员约纳斯·盖平表示，BDH-CQ模型是专门为ARC类问题构建的，因此很难与通用人工智能系统直接比较。但他称这种方法“很巧妙”，并指出它无需重新训练就能应对测试问题。

对邓而言，这项研究提出了一个更广泛的问题：人工智能是否需要将其所有推理都转化为文字？“人类语言对于交流推理过程是有用的，但它未必是每一次中间计算的最高效表示方式，”邓说。

英文来源：

Can AI reason without words? A small model puts the idea to the test
Skipping the words could make some AI reasoning more efficient
AI may be able to reason without putting every step into words.
Many of today’s AI systems tackle difficult problems by generating intermediate steps in lengthy, chains of words before arriving at an answer— a technique often called chain-of-thought reasoning. That approach can improve performance on some tasks, but it also means an AI model may generate many words, or tokens, on the way to a relatively short answer. Each token takes computing power to produce, so longer reasoning can make answers slower and more expensive to generate.
A small, experimental AI system called BDH-CQ takes a different approach. It can solve some reasoning puzzles without spelling out its intermediate thinking, researchers report in a paper submitted August 10 to arXiv.org. The work asks whether AI needs language at every step of reasoning, and whether doing more internally could make some reasoning cheaper.
In training, AI models are often shown a few examples before they try a new problem. Many keep those examples in front of them as they work. But on reasoning tests, BDH-CQ (DH is short for Dragon Hatchling) instead uses each example to update a fixed-size memory, which stays the same size as it takes in more examples, so it does not have to keep going back through everything it has already seen.
“Once the query arrives, the model doesn’t write out its thinking in words at all,” says complexity scientist Zuzanna Stamirowska, CEO of AI company Pathway. Stamirowska says the model works through a problem internally without turning each step into words. “Nothing in between ever converts into language,” she says.
Stamirowska and colleagues said the model solved nearly three in 10 puzzles on the public ARC-AGI-1 evaluation set when given two attempts. The result showed that a small model could solve some new reasoning problems without spelling out its intermediate steps. The benchmark tests ask AI systems to infer visual rules from a few examples and apply them to new puzzles. The tests are like human student exams.
The model did better on some puzzles than others. For example, it handled some tasks involving turning or moving shapes but struggled more with color changes and some combinations of rules. On harder ordering and nesting puzzles, the model did better after seeing an example of similar difficulty.
The process with fewer words may help keep computing costs down, Stamirowska and colleagues say. They estimate that each puzzle query costs about $0.00070 to run, about one-eleventh as much as GPT-5.6 Luna on the same benchmark test, though the two costs were calculated differently. The study has not been peer-reviewed.
The finding is “an interesting efficiency result,” says computer scientist Yuntian Deng of the University of Waterloo in Canada. But he says it does not show that BDH-CQ’s underlying architecture is better than other approaches.
“Specialized small models [like BDH-CQ] are a promising route to efficient reasoning,” Deng says. But more testing is needed to separate the effects of the model’s design from how it was trained, he says. Without written-out reasoning, the model is harder to inspect, Deng says. But even a written chain of thought may not faithfully show how the model reached its answer.
The BDH-CQ model was built specifically for ARC-style problems, which makes it hard to compare directly with general-purpose AI systems, says machine learning researcher Jonas Geiping of the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems in Germany. But he called the approach “neat” and noted that it can tackle test problems without retraining.
For Deng, the study raises a broader question: Does AI need to put all of its reasoning into words?. “Human language is useful for communicating reasoning, but it need not be the most efficient representation for every intermediate computation,” Deng says.