【Hacker News搬运】DeepSeek：通过大规模合成数据推进LLM中的定理证明

[hackernews]

Title: DeepSeek: Advancing theorem proving in LLMs through large-scale synthetic data

DeepSeek：通过大规模合成数据推进LLM中的定理证明

Text:

hn link

Url: https://arxiv.org/abs/2405.14333

由于我是一个文本模型，无法直接访问或处理外部链接。不过，我可以提供一些指导，说明如果你想要使用JinaReader或类似工具来抓取和分析论文内容，以及如何进行内容总结和翻译的步骤。

1. **抓取内容**：
   - 使用JinaReader或其他爬虫工具，你需要访问上述提供的arXiv链接。
   - 你可能需要编写一个脚本或使用JinaReader提供的API来下载PDF文件或HTML页面。

2. **分析内容**：
   - 一旦你有了论文的内容，你可以使用JinaReader的文本分析功能来提取关键信息。
   - 这可能包括提取标题、作者、摘要、关键词、段落内容等。

3. **总结内容**：
   - 对于提取的关键信息，你可以编写一个简单的算法或使用JinaReader提供的功能来生成摘要。
   - 这通常涉及对文档的每个部分进行简短的重述，并合并这些重述以形成一个连贯的摘要。

4. **翻译内容**：
   - 如果内容不是中文，你可以使用在线翻译服务或集成翻译API（如Google Translate API）来翻译文本。
   - 你需要编写代码来调用翻译API，并将抓取的文本传递给它。
   - 翻译后的文本可以用来进行总结，或者你可以直接阅读翻译后的内容。

以下是一个简化的示例流程，展示了如何使用伪代码来实现上述步骤：

```python
# 伪代码示例

# 步骤1: 抓取内容
pdf_content = fetch_pdf_content("https://arxiv.org/abs/2405.14333")

# 步骤2: 分析内容
key_info = analyze_content(pdf_content)

# 步骤3: 总结内容
summary = generate_summary(key_info)

# 步骤4: 翻译内容
translated_summary = translate_text(summary, target_language="zh")

# 输出翻译后的总结
print(translated_summary)

# 以下为伪代码函数定义
def fetch_pdf_content(url):
    # 实现抓取PDF内容的逻辑
    pass

def analyze_content(content):
    # 实现分析内容的逻辑，如提取关键信息
    pass

def generate_summary(key_info):
    # 实现生成摘要的逻辑
    pass

def translate_text(text, target_language):
    # 实现调用翻译API的逻辑
    pass

请注意，上述代码仅为伪代码，并不是实际的实现。在实际应用中，你需要根据JinaReader和翻译API的具体文档来实现这些功能。

        
## Post by: hhs
        
### Comments: 
        
**aabhay**: The ability to use automatic verification + synthetic data is basically common knowledge among practitioners. But all these organizations have also explored endlessly the different ways to overfit on such data and the conclusion is the same -- the current model architecture seems to plateau when it comes to multi-step logical reasoning. You either drift from your common knowledge pre-training too far or you never come up with the right steps in instances where there&#x27;s a vast design space.<p>Think -- why has nobody been able to make an LLM play Go better than AlphaZero while still retaining language capabilities? It certainly would have orders of magnitude more parameters.
> **aabhay**: 使用自动验证+合成数据的能力基本上是从业者的常识。但所有这些组织也在不断探索对这些数据进行过拟合的不同方法，结论是一样的——当涉及到多步逻辑推理时，当前的模型架构似乎停滞不前。你要么偏离了训练前的常识太远，要么在以下情况下从未提出正确的步骤：；这是一个广阔的设计空间<p> 想想看——为什么没有人能够在保持语言能力的同时，让LLM玩围棋比AlphaZero更好？它肯定会有更多数量级的参数。
            
**maxrmk**: There&#x27;s a newer version of this model that takes a really cool RL based approach: <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;pdf&#x2F;2408.08152" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;pdf&#x2F;2408.08152</a>
> **maxrmk**: 那里；这是此模型的一个较新版本，它采用了一种非常酷的基于RL的方法：<a href=“https:&#x2F；arxiv.org&#x20F；pdf&#x2048152”rel=“nofollow”>https:&#x2F；arxiv.org；pdf；2408.08152</a>
            
**_flux**: This must be one of the best applications for LLMs, as you can always automatically verify the results, or reject them otherwise, right?
> **_flux**: 这一定是LLM的最佳应用程序之一，因为您始终可以自动验证结果，否则可以拒绝它们，对吧？
            
**youoy**: From the abstract:<p>&gt; Proof assistants like Lean have revolutionized mathematical proof verification... 
&gt; Both the synthetic dataset and the model will be made available to facilitate further research in this promising field.<p>Please don&#x27;t use &quot;revolutionised&quot;, &quot;promising&quot; on a scientific abstract... If it has revolutionised, tell me instead what was the important thing that made it happen. If it is promising, tell me instead why.<p>Appart from that, the second sentence says:<p>&gt; Although large language models (LLMs) show promise in mathematical reasoning, their advancement in formal theorem proving is hindered by a lack of training data.<p>Is it really? Of hindered by unsuitable model architecture? Because you could perefectly say that Lean is an AI capable of theorem proving... I feel that that sentence should be the potential conclusion of the paper, not the first observation. Like, &quot;we used more training data and it works better, if we scale theorem proving training data, we can have the performance that we want, so training data is <i>actually</i> hindering the advancement in formal theorem proving on LLMs&quot;.
> **youoy**: 摘要：<p>&gt；像Lean这样的证明助手彻底改变了数学证明验证。。。&gt；合成数据集和模型都将用于促进这一有前景的领域的进一步研究<p> 请不要；不要使用&quot；革命”&“；有希望&quot；在一篇科学摘要中。。。如果它发生了革命性的变化，请告诉我是什么让它发生的重要事情。如果有希望，告诉我为什么<p> 从那以后，第二句话说：<p>&gt；尽管大型语言模型（LLMs）在数学推理方面显示出希望，但由于缺乏训练数据，它们在形式定理证明方面的进步受到了阻碍<p> 真的吗？是否受到不合适的模型架构的阻碍？因为你可以完美地说精益是一种能够证明定理的人工智能。。。我认为这句话应该是论文的潜在结论，而不是第一个观察。例如，&quot；我们使用了更多的训练数据，效果更好，如果我们缩放定理证明训练数据，我们可以获得我们想要的性能，因此训练数据实际上阻碍了LLM上形式定理证明的进步&quot；。
            
**whyowhy3484939**: &quot;Suppose you try to construct a coherent, ordered, natural world with no resource other than repeated exposure to things, and the formation of certain associative bonds. Oh, please!&quot;<p>This is prof. Robinson on Kantian philosophy - check out Oxford podcasts by the way - and this quote is meant to imply that building a coherent world out of raw sensory data and statistics alone is completely and utterly impractical if not outright impossible. While I don&#x27;t think he meant to refer to any kind of AI, in my mind this description also aptly describes the general method of DL neural networks. Repeated exposure to find correlation.<p>How does one find order through associativity alone? With AI this is not an academic problem anymore. This has become practical. Kant says it is impossible, not just unlikely.<p>The Kantian project and the various core issues it tries to address seems readily applicable to AI research yet I see very little mention of it. Perhaps I am just dumb though. Building a mind capable of taming tremendous sensory flux needs to, at the very least, take note of the (many) fundamental issues he raised. Issues I feel are not at all trivial to set aside. I feel we are stuck in Hume&#x27;s empiricist reasoning and have yet to graduate to Kant and beyond.<p>Are we now somehow convinced yet again that causality and reasoning will, in fact, after all spontaneously emerge out of pure chaos? Didn&#x27;t we settle the impossibility of this a few hundred years ago?
> **whyowhy3484939**: &“；假设你试图构建一个连贯、有序、自然的世界，除了反复接触事物和形成某些关联关系外，没有其他资源。拜托&“<p> 这是罗宾逊教授对康德哲学的看法——顺便看看牛津播客——这句话的意思是，仅凭原始感官数据和统计数据建立一个连贯的世界是完全不切实际的，如果不是完全不可能的话。虽然我不知道；我不认为他指的是任何一种人工智能，在我看来，这种描述也恰当地描述了DL神经网络的一般方法。反复接触以寻找相关性<p> 如何仅通过结合性找到秩序？对于AI来说，这不再是一个学术问题。这已经变得切实可行。康德说这是不可能的，而不仅仅是不可能<p> 康德项目及其试图解决的各种核心问题似乎很容易适用于人工智能研究，但我很少看到有人提到它。也许我只是很愚蠢。建立一个能够驯服巨大感官通量的头脑，至少需要注意他提出的（许多）基本问题。我觉得这些问题根本不是微不足道的。我觉得我们被困在休谟了；他的经验主义推理尚未发展到康德及以后<p> 我们现在是否再次确信因果关系和推理最终会自发地从纯粹的混乱中出现？没有；我们不是在几百年前就解决了这个不可能的问题吗？