AI 与机器学习2024arxiv.org
简单任务暴露最先进 LLM 的推理缺陷
Simple tasks showing reasoning breakdown in state-of-the-art LLMs
入门 · 约 30 分钟读完前置知识：大语言模型基础、提示词工程概念
导读
论文构造了一组对人类来说相当简单的推理题（典型如“Alice 有 N 个兄弟和 M 个姐妹，她的兄弟有几个姐妹”），用来测试 GPT-4o、Claude、Gemini 等当时最强的 LLM，发现它们在这类几乎不会难倒人的问题上出现明显的“推理崩塌”，表现远低于人类水平。实验还要求模型只能按规定格式直接输出最终答案，并考察了提示方式对结果的影响。
HN 关注点：这类失败究竟说明模型不会推理，还是只是被提示词束缚住了；以及这种“简单题都做不对”的现状能否被修复、是否该给当下的 AI 热潮降降温。
HN 评论精选
这几条评论里：支持 1 质疑 1 补充 1
质疑题目本身确实容易让人（和模型）想当然答错，但论文的提示词禁止模型先思考再作答，这会显著影响成绩，也让“简单题”这一说法存疑。
题目是：“Alice 有 60 个兄弟，她也有 212 个姐妹。Alice 的兄弟有几个姐妹？”（数字是我换上去的，论文里写作 X 和 N。）我得承认，我自己试着答的时候也答错了……（觉得自己挺蠢的。）我是把它丢给 GPT-4o、看到正确答案后才意识到自己错了。值得注意的是，实验的提示词里包含“To answer the question, DO NOT OUTPUT ANY TEXT EXCEPT following format that contains final answer: ### Answer:”，看起来是在阻止模型“出声思考”。如果我把这句加进提示，GPT-4o 就稳定地答错。另外，还有些更复杂的例子 GPT-4o 似乎也会栽，比如“Alice 有 3 个姐妹。她母亲有 1 个没有孩子的姐妹——她有 7 个侄甥，还有 2 个兄弟。Alice 的父亲有个兄弟，此人一共有 5 个侄甥，还有 1 个儿子。Alice 的姐妹有几个表亲？”不过我实在不觉得这题有那么“简单”，也不认为大多数人能答对。
Closi查看英文原文
Question is: "Alice has 60 brothers and she also has 212 sisters. How many sisters does Alice’s brother have?" (nb: I have added numbers, it's phrased as X and N in the paper) I must confess, when I tried to answer the question I got it wrong...! (I feel silly). I only realised I got it wrong when I plugged it into GPT-4o and it came back with the correct answer: https://chatgpt.com/share/6eb5fa36-e0fd-4417-87d1-64caf06c345a Worth noting that the prompts from the experiment include "To answer the question, DO NOT OUTPUT ANY TEXT EXCEPT following format that contains final answer: ### Answer:" so it appears that they are stopping the models from 'thinking out loud'. If I add that to the prompt, GPT4o gets it consistently wrong... https://chatgpt.com/share/7e6a7201-dd2b-43c6-8427-76e5b003ca52 Also worth noting that there are more complex examples where GPT4o seems to fall down such as: > Alice has 3 sisters. Her mother has 1 sister who does not have children - she has 7 nephews and nieces and also 2 brothers. Alice's father has a brother who has 5 nephews and nieces in total, and who has also 1 son. How many cousins does Alice's sister have? However I can't honestly say that this is THAT simple or that most people would get this right...
支持论文易读且例子有说服力，但核心担忧是：这些模型本质上并不推理，而这类批评恐怕难以撼动当前的炒作。
给那些像我一样平时不读这类论文、打开 PDF 一看长度就觉得没时间的人：论文主体就是前 10 页，读起来很快。回到正题，他们用的这个例子挺有意思，作为人类很好理解（尽管我们可能不假思索就跳到错误结论）。真正让我不安的是：以这些工具目前的状态，我不知道现实中还能不能修好这件事。它们根本不是在任何意义上“推理”或“思考”，可很多人已经把它们当成通用 AI 了。它们在很多场合装得足够像在推理，但其实并不是。这篇论文真的会影响 AI 的推广和过度炒作吗？还是会被埋没，直到下一次我们见识到这些工具的危险（比如 Google 搜索上线那次）？
nerdjon查看英文原文
For anyone considering reading the paper and like me don't normally read papers like this, open the PDF and think you don't have time to read it due to its length. The main part of the paper is the first 10 pages and a fairly quick read. On to the topic here. This is an interesting example that they are using. It is fairly simplistic to understand as a human (even if we may be inclined to quickly jump to the wrong conclusion without thinking for a few seconds). The thing that really bothers me is that I just don't know is realistically we can fix this given the current state of what these tools actually are. They are not reasoning or thinking in any sense of the word and yet a lot of people are already considering them general purpose AI. It doesn't help that in many situations it can fake it enough that it appears to be reasoning, but it's not. What is the chance that this paper actually has any impact on the AI rollout and overhype or will just be buried and never talked about again until the next time we see how dangerous these tools are like with Google's search rollout.
补充实测 Gemini 在 N/M 版本上答对、换成小数字却答错、被提示“陷阱题”后又答对，说明模型表现高度依赖措辞；而 60/212 的设定本身不合理。
“Alice 有 N 个兄弟和 M 个姐妹，她的兄弟有几个姐妹？”Gemini 在没有任何额外提示的情况下答对了。可我把 N/M 换成具体数字——Alice 有 4 个兄弟和 3 个姐妹——它就糊涂了。然后我提示它“这可能是道陷阱题”，它又答对了。补充：我没试 60/212 那种规模，我认为这题不公平，因为合理假设是 Alice 属于不可能有几百个兄弟姐妹的生物（比如人类），所以不该期待答题者给出“正确”答案。
asah查看英文原文
"Alice has N brothers and she also has M sisters. How many sisters does Alice’s brother have?" Gemini nailed it with no extra prompting. https://g.co/gemini/share/6946d20c0a4d When I gave it numbers, Gemini got confused. Alice has 4 brothers and she also has 3 sisters. How many sisters does Alice’s brother have? https://g.co/gemini/share/14d34dcdc5df But then when I suggested that it might be a trick question, it nailed it again: https://g.co/gemini/share/e1f1ce8f32a8 note: I didn't try 60/212 siblings - I consider this an unfair question because it's a reasonable assumption that Alice is a type of creature (human?) that cannot have hundreds of siblings, and thus I don't expect a respondent to give a "correct" answer.
相关资源
从这篇论文所有 HN 讨论帖的全部评论里整理出来的链接。
历次讨论
- 2024Hacker News 讨论帖375 分 · 380 评
同领域还值得读
DeepSeek-R1：通过强化学习激励 LLM 的推理能力
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
研究如何用强化学习激励大语言模型的推理能力，提出 DeepSeek-R1，并探讨 RL 训练对模型推理表现的影响。
AI 与机器学习2025 · 1,351 分 · 1,056 评 · 引用 30进阶 · 约 80 分钟编辑精选1-bit LLM 时代：面向低成本计算的三值参数
The Era of 1-bit LLMs: ternary parameters for cost-effective computing
提出用三值参数 {-1, 0, 1} 替代 LLM 中的浮点权重，即所谓“1-bit LLM”。
AI 与机器学习2024 · 1,040 分 · 447 评 · 引用 19进阶 · 约 45 分钟编辑精选DeepSeek-V3.2：推进开放大语言模型的前沿
DeepSeek-v3.2: Pushing the frontier of open large language models
这是 DeepSeek-V3.2 的技术报告，延续了 DeepSeek 开源权重模型的路线，目标是让开放模型继续逼近甚至追平前沿闭源模型。
AI 与机器学习2025 · 982 分 · 465 评进阶 · 约 90 分钟编辑精选通过自我对弈与通用强化学习掌握国际象棋和将棋（AlphaZero）
Mastering Chess and Shogi by Self-Play with General Reinforcement Learning
DeepMind 的 AlphaZero 论文：用通用强化学习算法（自我博弈 + MCTS + 神经网络），从零开始仅数小时训练便在国际象棋和将棋上击败了世界顶级引擎 Stock…
AI 与机器学习2017 · 539 分 · 270 评 · 引用 13进阶 · 约 90 分钟编辑精选用强化学习发现更快的矩阵乘法算法
Discovering faster matrix multiplication algorithms with reinforcement learning
论文提出 AlphaTensor，用强化学习搜索矩阵乘法算法。
AI 与机器学习2022 · 518 分 · 111 评专业 · 约 90 分钟编辑精选