在近期，一项旨在测试大模型数学理解能力的实验引发广泛关注。这一实验以简单的数学问题“9.11和9.9哪个大？”作为测试题，却发现即使是先进的大模型，也无法准确解答。这一现象不仅揭示了当前人工智能在处理数学问题时的局限性，也引发了对大模型智能水平的深入思考。
实验由知名人工智能研究机构Scale AI发起，其高级提示工程师Riley Goodside在测试中发现，当提出“9.11和9.9哪个大？”的问题时，大模型GPT-4o给出了错误答案，声称9.11大于9.9。随后，实验团队对包括GPT-4o、Claude-3.5-Sonnet、以及谷歌的Gemini在内的15款大模型进行了实测，结果发现超过半数的大模型在这一基本数学问题上“翻车”，未能正确识别数值大小。
为了更深入地了解这一现象，实验团队选取了12款国内大模型以及国外的GPT-4o、Claude-3.5-Sonnet和谷歌的Gemini进行了集中评测。结果显示，大模型在处理这一问题时普遍表现出不足，尤其是在处理小数部分的比较时，出现了明显的错误。这一测试不仅反映了当前大模型在数学逻辑理解方面的局限性，同时也提示我们，尽管人工智能在许多领域展现出强大的能力，但在特定的、基础的数学理解与处理上，仍存在明显的挑战。
此次实验的结果引发了广泛的讨论，对于人工智能领域的发展具有重要意义。它不仅提醒我们人工智能在数学理解上的不足，也为未来的研究和开发指明了方向，即如何提升大模型在基础数学知识理解与应用上的能力，以期在更广泛的领域内实现更高效、更准确的智能服务。
英语如下：
News Title: “Major AI Model Mishap: 9.11 vs 9.9, Who’s the Winner?”
Keywords: Major AI Model Mishap, Mathematical Challenge, 9.11 vs 9.9
News Content: In recent times, an experiment aimed at gauging the mathematical comprehension ability of advanced AI models has garnered significant attention. This experiment posed a straightforward mathematical question, “Which is larger, 9.11 or 9.9?” as its test, yet even sophisticated AI models were found unable to provide an accurate answer. This phenomenon not only highlights the current limitations of artificial intelligence in handling mathematical problems but also prompts a deeper contemplation on the intelligence levels of AI models.
Initiated by the renowned AI research institution Scale AI, the experiment, led by senior prompt engineer Riley Goodside, discovered that when asked the question, “Which is larger, 9.11 or 9.9?”, the AI model GPT-4o incorrectly stated that 9.11 is greater than 9.9. Subsequently, a team tested 15 AI models, including GPT-4o, Claude-3.5-Sonnet, and Google’s Gemini, and found that more than half of these models “flunked” the basic math question, failing to correctly identify the numerical size.
To delve deeper into this phenomenon, the research team selected 12 domestic AI models alongside GPT-4o, Claude-3.5-Sonnet, and Google’s Gemini for a comprehensive evaluation. The results revealed that AI models generally demonstrated deficiencies in dealing with this problem, particularly in comparing decimal parts, leading to clear errors. This test not only reflects the current limitations of AI models in understanding and applying mathematical logic but also signals that, despite the impressive capabilities of artificial intelligence in many fields, there are evident challenges in specific, foundational areas of mathematical understanding and processing.
The outcomes of this experiment have sparked wide-ranging discussions, holding significant importance for the development of the AI field. It not only underscores the inadequacies of AI in mathematics but also points the way forward for future research and development, emphasizing the need to enhance AI models’ ability to understand and apply basic mathematical knowledge in a broader range of applications to achieve more efficient and accurate intelligent services.
【来源】https://www.jiqizhixin.com/articles/2024-07-18-4
Views: 2