AI 与机器学习2024arxiv.org
推理会涌现吗？大语言模型中的因果概率
Does Reasoning Emerge? Probabilities of Causation in Large Language Models
进阶 · 约 60 分钟读完前置知识：因果推断基础、概率论、LLM 基础
导读
论文提出用因果推理中的必要性概率（PN）和充分性概率（PS）来衡量 LLM 是否具备“真正的推理”能力：通过设计特定的事实性与反事实性测试题，测量模型在因果链推断上的错误率。结果显示 LLM 对反事实错误的容忍度出奇地低，作者认为这可能表明当前 LLM 架构的推理能力存在上限。
HN 关注点：HN 读者长期争论 LLM 到底是“真推理”还是“模式匹配”，这篇论文提供了用因果概率量化这一问题的框架，讨论焦点是其方法是否可靠以及结论是否意味着当前架构的瓶颈。
HN 评论精选
这几条评论里：支持 2 质疑 1 补充 1
支持LLM 只是低层次模式匹配，无法真正符号化抽象，推理表现严重依赖提问内容与形式。
我的印象是，LLM 的“模式匹配”抽象层次低于通用推理所需。它们通过训练掌握了大量典型推理模式，但这些模式并未与训练数据中具体实例的内容充分解耦或泛化。因此 LLM 表现出的推理能力很大程度上取决于被问的内容，甚至取决于提问的句式。LLM 似乎基本无法从训练数据示例中进行符号化抽象（而非插值）。
layer8查看英文原文
My impression is that LLMs “pattern-match” on a less abstract level than general-purpose reasoning requires. They capture a large number of typical reasoning patterns through their training, but it is not sufficiently decoupled, or generalized, from what the reasoning is about in each of the concrete instances that occur in the training data. As a result, the apparent reasoning capability that LLMs exhibit significantly depends on what they are asked to reason about, and even depends on representational aspects like the sentence patterns used in the query. LLMs seem to be largely unable to symbolically abstract (as opposed to interpolate) from what is exemplified in the training data.
支持推理是语义之外的能力，LLM 只是复现语料中的推理痕迹，不应期待其涌现。
LLM 掌握的是人类集体语义理解的空间。我不明白为什么人们期待明显属于语义之外（extra-semantic）的认知能力会从中自然涌现。它们有时看起来在推理，只是因为人类文本语料中包含大量推理内容——但那只是非语义过程产生的语义副产品。人类认知远不止把句子串起来。
heyjamesknight查看英文原文
LLMs have access to the space of collective semantic understanding. I don't understand why people expect cognitive faculties that are clearly extra-semantic to just fall out of them eventually. The reason they sometimes appear to reason is because there's a lot of reasoning in the corpus of human text activity. But that's just a semantic artifact of a non-semantic process. Human cognition is much more than just our ability to string sentences together.
质疑质疑：这类基准一旦流行就会被模型针对性训练过拟合，无法持续衡量真实推理能力。
论文提出用某种“陷阱问题”来测量“真正的推理”，但我不认为这能成为可持续的基准。一旦受到关注，下一代 LLM 就会拿这篇论文来训练，再用这类问题做微调以在该基准上表现良好……然后我们又回到了原点。
doe_eyes查看英文原文
This is proposed as a way to measure "true" reasoning by asking a certain type of trick questions, but I don't quite see how this could be a basis of a sustainable benchmark. If this gets attention, the next generation of LLMs will be trained on this paper, and then fine-tuned by using this exact form of questions to appear strong on this benchmark, and... we're back to square one.
补充肯定 PN/PS 框架的价值，但指出其局限于简单布尔链，且低反事实容忍度可能预示架构上限。
他们的假设不错：推理的一种形式是通过 PN 和 PS 连接因果；可以基于 LLM 建模判断自然语言输入能否支持 PN/PS 推断。这意味着可以通过数据和架构设计注入更多因果推理。他们定义了因果函数，把准确率（假阳/假阴）投影到事实与反事实断言测试上，发现对反事实错误率的容忍度出奇地低，可能暗示当前架构推理的上限。但方法受限于简短的布尔因果链，难以扩展到复杂推理；反过来说，如果 LLM 连这都做不好，除非方法遗漏了更大模型的某种动态，否则很难按预期速度进步。这至少可作为简单推理的 LLM 质量工程的起点。
w10-1查看英文原文
Their hypothesis is a good one: - A form of reasoning is to connect cause and effect via probability of necessity (PN) and the probability of sufficiency (PS). - You can identify when the natural language inputs can support PN and PS inference based on LLM modeling That would mean you can engineer in more causal reasoning based on data input and model architecture. They define causal functions, project accuracy measures (false positives/negatives) onto factual and counter-factual assertion tests, and measure LLM performance wrt this accuracy. They establish surprisingly low tolerance for counterfactual error rate, and suggest it might indicate an upper limit for reasoning based on current LLM architectures. Their findings are limited by how constrained their approach is (short simple boolean chains). It's hard to see how this approach could be extended to more complex reasoning. Conversely, if/since LLM's can't get this right, it's hard to see them progressing at the rates hoped, unless this approach somehow misses a dynamic of a larger model. It seems like this would be a very useful starting point for LLM quality engineering, at least for simple inference.
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165 分 · 192 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