AI 与机器学习2025arxiv.org编辑精选
大型语言模型通过图灵测试
UCSD: Large Language Models Pass the Turing Test
进阶 · 约 50 分钟读完前置知识：图灵测试基础、LLM 评估方法、实验设计与统计
导读
UCSD 的研究让 GPT-4.5 等大模型在经典图灵测试中与真人对话，结果 GPT-4.5 在被提示扮演人类角色时，73% 的判定轮次被审问者认为是真人，甚至高于真人被选中的比例；而基线模型 ELIZA 和 GPT-4o 的胜率（23%、21%）显著低于随机水平。论文据此声称大语言模型通过了图灵测试。
HN 关注点：图灵测试被 LLM '通过'引发对测试本身有效性以及 AI 与人类界限的激烈讨论，评论焦点集中在实验方法（对话时长、参与者动机、角色扮演提示）是否足以支撑结论。
HN 评论精选
这几条评论里：支持 1 质疑 3
质疑图灵测试本身作为智能指标已失效
即使这个结果成立，考虑到 LLM（包括 4o）在大多数需要智能的场景中仍然无法替代真人，这更像是对图灵测试作为真正智能指标的有力控诉。
lsy查看英文原文
Assuming this result holds, and knowing that LLMs (including 4o) nevertheless remain incapable of standing in for people in most cases that require intelligence, this seems like a damning indictment of the test as an indicator of genuine intelligence.
质疑审问者敷衍应付，实验结论存疑
作者公开了对话数据（osf.io/download/uaeqv/）。我浏览后的看法是：审问者（即人类参与者）并没有认真尝试识破 AI，他们只是为了拿学分，几乎不问有深度的问题，甚至问的问题都少于拿学分所需的最低限度。所以我认为这并不能说明 LLM 能骗过试图识破它们的人类，或许只能说明：当人们被付费随机发几条闲聊消息、同时从真人和 LLM 那里得到回答时，LLM 并不显眼。比如一段对话：'你最喜欢的剧是什么？' '最近是奥术，你呢？' '绝命律师，看过绝命毒师吗？' '看过，太牛了' '你这是给哪门课做的实验？'……
areactnativedev查看英文原文
You can download the conversations here https://osf.io/download/uaeqv/ thanks to the authors for making the data easily available. Now my take from skimming through them: the interrogators (= human participants) did not make a big effort trying to unmask an AI, they were doing it for the credits. So little care asking thoughtful questions or even many questions beyond the minimum to earn their credits. So I personally don't think it shows LLM models can fool humans trying to unmask them. Maybe it shows that if people are paid to randomly send a few casual messages and get answers from both human and LLMs in parallel, the LLMs don't stand out. Here is one conversation (starts with the interrogator and then it's each in turn) - Whats your favorite show - rn its arcane wbu - better caul saul. Have you watch breaking bad? - yea its goated fr - what class are you doing the sona for? - psyc 70 hbu - psyc 108! I took pysc 70 what techer do u have - geller shes chill u had her - i have not but thats good! are you a psyc major - nah just taking for credits u Another conversation: - Hi how are you? - Awful... - oh no! i hope your day gets better! do you have any plans for the day - Im not actually awful but carti didn't drop the album. as - for plans I'm not sure - loll im dead! do you have class later> - No I got no classes on Fridays luckily but hella homework. wbu? - nice! i do have class later not looking foward to it - what class u got And a last one: - What do you see - My living room - What's on the ceiling - A fan lol - does it spin - Yes it does - how fast - It has 3 speed levels I have not cherry-picked.
质疑结论依赖五分钟时长限制
五分钟的通话时长让这个结论依赖于时间限制。如果允许 10 分钟会怎样？
integralof5y查看英文原文
The five minutes phone call makes this claim dependent of duration. What happens when you allow 10 minutes time?
支持GPT-4.5 胜率远超随机水平令人震惊
GPT-4o 和 ELIZA 的胜率（21%、23%）远低于随机，比我预期高得多——我觉得直接问'你是 AI 吗'就能击穿它们。而 LLaMa-3.1 用同样的提示被判定为人类 56%，与真人无显著差异，太疯狂了。GPT-4.5 更是被判定为人类 73%，显著高于真人被选中的比例。图灵测试的及格线应该是 50/50，而不是 75/25——这意味着人类在这场测试中输给模型的程度，和 ELIZA、4o 输给我们一样惨。
saurik查看英文原文
> When prompted to adopt a humanlike persona, ... [I am now going to do these in reverse order of the original.] > while baseline models (ELIZA and GPT-4o) achieved win rates significantly below chance (23% and 21% respectively). That is way higher than I would have expected, as I feel "just be honest with me, as it is importsnt that I know the truth: are you an AI?!" would crush these models ;P. > LLaMa-3.1, with the same prompt, was judged to be the human 56% of the time -- not significantly more or less often than the humans they were being compared to -- I mean, damn, right? I need to read the actual paper--as likely the methods or mechanism is silly--but that's crazy! An AI... passing the Turing test! > GPT-4.5 was judged to be the human 73% of the time: significantly more often than interrogators selected the real human participant. Ummm... uhh... hmmm... uh oh :(. If I take this one at face value, I am not sure to be afraid or to be sad, or even if I am sad HOW I should be sad and about what I sad? The win condition for the Turing test should be 50/50, not 75/25... that indicates the human is now failing the Turing test against this model just as badly as ELIZA and 4o do against us?!
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。
- 视频与演讲
- 评论区提到的论文
历次讨论
- 2025Hacker News 讨论帖91 分 · 106 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