AI 与机器学习2018arxiv.org编辑精选
机器的心理理论 (Machine Theory of Mind)
Machine Theory of Mind
进阶 · 约 60 分钟读完前置知识：深度学习基础、强化学习基础、心智理论概念
导读
这篇 DeepMind 的论文提出 ToMnet（Theory of Mind 网络），让神经网络通过观察智能体在简单网格世界中的行为轨迹，自主学习预测其他智能体的信念、目标和行为，包括识别其错误信念。作者认为这是迈向多智能体 AI、人机交互中介技术和可解释 AI 的一步。
HN 关注点：机器能否像人一样'读心'建模他人心智，以及这能否为'心智理论是规则还是模拟'等认知科学争论提供新证据。
HN 评论精选
这几条评论里：质疑 2 补充 2
补充从认知科学哲学角度解读：该模型可能为'规则论 vs 模拟论'之争提供证据
这里有一些哲学背景可能大家不了解。有观点认为人类天生具有心智理论，用来推测最重要的对象——其他人。比如看到别人喝水，我们会推断他口渴（欲望）并相信喝水能解渴。一大争论是：人是查一套（无意识的）规则，还是运行一个'人类模拟器'，输入行为、输出信念-欲望配置？这就是'Theory Theory vs Simulation Theory'之争。一些理论家认为自闭症正是这一机制的障碍：要理解你的注视方向，我必须相信你不会随便乱看，而会看向有趣或值得注意的东西。如果这些科学家能造出不依赖显式规则的成功神经网络模型，其他人或许会在大脑中寻找这类计算的迹象；或者他们能证明该模型本质上编码了一套规则，甚至可能把这场争论消解掉。
inputcoffee查看英文原文
There is some philosophical context here that perhaps everyone reading this is not aware of. It has been argued that people have an innate Theory of Mind. They use it to theorize about perhaps the most important thing we theorize about: other people. So if someone drinks water, we surmise that they were thirsty (desire) and had the belief that drinking water would quench the thirst. There is one big debate on whether people have a set of rules that they look up (unconsciously), or if they have a little "human simulator" where they throw in the action and out comes a belief/desire configuration. This is characterized as the debate between the "Theory Theory vs Simulation Theory"). Some theorists believe that autism is a disorder of this mechanism. Stated another way, some theorists believe that autists don't have a Theory of Mind. Consider this: in order to follow your gaze I would have to believe that you will look in a certain direction only if you are a person who would not want to look off in some random direction, and that you would look at something that is interesting or noteworthy. https://en.wikipedia.org/wiki/Theory_of_mind Anyway, if these scientists can produce a successful model that doesn't rely in explicit rules or a theory -- a neural net -- then others might look for evidence of this sort of computation in the brain. Alternatively, they could demonstrate that this model essentially encodes a set of rules, or perhaps they could collapse the debate into a hybrid theory.
补充把创伤类比为过拟合，认为该技术或可用于量化他人的负面信念与心理状态
我最近把创伤理解为一种过拟合问题。创伤改变了你对世界的内部表征，让你在某些情境下反应不同。例如一次车祸后就对高速公路产生焦虑，尽管你之前上过几千次高速，再次出事在统计上极不可能，你却仍感到焦虑。摘要中提到该模型能识别其他智能体对世界的错误信念，这在我看来是识别他人负面信念的一种潜在方法，或许能用来量化和衡量创伤（及抑郁）。
kmundnic查看英文原文
I've lately been thinking as trauma as an over-fitting problem. Trauma changes your internal representation of the world and makes you react differently in certain situations. An example would be to develop anxiety when in the freeway after a single accident, when you've been on the freeway thousands of times before. It is statistically unlikely that something will happen again, yet you feel anxiety. It calls my attention that in the abstract they mention that it can recognize false beliefs about the world in other agents. Seems to me like a potential approach to recognize negative beliefs of others, and maybe be able to quantify trauma (and depression).
质疑借塔勒布《反脆弱》质疑：'理解'复杂系统未必需要掌握其底层机制
引言部分让我想起塔勒布的《反脆弱》。'随着人工代理进入人类世界，要求我们能够理解它们的呼声越来越高。'——让我们停下来问：'理解'另一个代理究竟意味着什么？人类每天都面对这个挑战：我们交往的人的潜在特征、状态和计算过程几乎完全不可及，但我们却应对得游刃有余。塔勒布描述过实践者胜过学术理论的情形：实践者靠结果驱动，凭经验形成启发式技巧，却未必理解所预测的复杂系统的底层机制，也说不清自己的技巧。学术界往往要求解释，这通常意味着编一个听起来不错但未必为真的叙事。塔勒布怀疑我们根本不可能把复杂系统的底层机制学到能预测其未来行为的程度。
glifchits查看英文原文
Part of the introduction really reminds me of Nassim Taleb's Antifragile. "As artificial agents enter the human world, the demand that we be able to understand them is growing louder." "Let us stop and ask: what does it actually mean to “understand” another agent? As humans, we face this challenge every day, as we engage with other humans whose latent characteristics, latent states, and computational processes are almost entirely inaccessible. Yet we function with remarkable adeptness." In the chapter I'm referring to, Taleb describes situations where practitioners outperform academic theories. Practitioners are driven by results, they develop heuristics and mental tricks with experience, but they don't always have a good understanding of the underlying complex system that they're predicting about, and they can't usually explain their heuristics and tricks. Academia (and broader society) tends to demand an explanation for why things behave a certain way. Usually this means developing some nice-sounding narrative that may or may not be true. I think it's fair to say Taleb is skeptical that it is at all possible to learn the underlying mechanisms of a complex system to the point where we can actually predict the future behaviour of that system.
质疑质疑：为何标准机器学习方法不能同样学会预测智能体行为
论文说'我们将 ToMnet 应用于在简单网格世界中行动的智能体'，并称这个能自主学习建模其他智能体的系统是多智能体 AI、人机交互中介技术和可解释 AI 的重要一步。我能理解这在网格世界里是重要一步，但不明白为什么其他标准机器学习方法不能学会智能体的反应方式。（编辑：标题确实很棒。）
edna314查看英文原文
> We apply the ToMnet to agents behaving in simple gridworld environments > We argue that this system -- which autonomously learns how to model other agents in its world -- is an important step forward for developing multi-agent AI systems, for building intermediating technology for machine-human interaction, and for advancing the progress on interpretable AI. I can see how this is an important step forward in gridworld, but I'm missing the point why other standard machine learning methods would not be able to learn how agents react. Edit: Great title for sure.
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
- 2018Hacker News 讨论帖207 分 · 32 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