<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Zain Mothupi on Medium]]></title>
        <description><![CDATA[Stories by Zain Mothupi on Medium]]></description>
        <link>https://medium.com/@Zain_Mothupi?source=rss-296ca9432f5d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Stories by Zain Mothupi on Medium</title>
            <link>https://medium.com/@Zain_Mothupi?source=rss-296ca9432f5d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 02:27:38 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@Zain_Mothupi/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Using language models to prove truths about reality]]></title>
            <link>https://medium.com/warwick-artificial-intelligence/using-language-models-to-prove-truths-about-reality-cb589055685b?source=rss-296ca9432f5d------2</link>
            <guid isPermaLink="false">https://medium.com/p/cb589055685b</guid>
            <category><![CDATA[gpt-3]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[theorem-prover]]></category>
            <dc:creator><![CDATA[Zain Mothupi]]></dc:creator>
            <pubDate>Mon, 17 Jan 2022 20:44:07 GMT</pubDate>
            <atom:updated>2022-01-17T20:44:07.103Z</atom:updated>
            <content:encoded><![CDATA[<h3>A talk by Stanislas Polu from OpenAI</h3><p>See more from Stanislas here: <a href="https://www.linkedin.com/in/spolu?lipi=urn%3Ali%3Apage%3Ad_flagship3_profile_view_base_contact_details%3BxmAZJiZjQ5CYIPfl0FN0Hg%3D%3D">linkedin.com/in/spolu</a></p><p>On 11th January, Warwick AI hosted a talk by Stanislas Polu on ‘Generative Language Modelling for Automated Theorem Proving’. Stanislas Polu is a research engineer at Open AI working on generative language modelling. He received his master’s degree in Computer Science from Stanford and has previously worked at top tier software engineering companies including Stripe and Oracle. He is also the co-founder of TOTEMS, a marketing company that uses data analytics to help customers advertise their content.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1IJw_MwI6YlnhzAv.png" /></figure><p><strong>Motivation</strong></p><p>Stanislas began by motivating Open AI’s decision to apply transformer-based language models to formal mathematics. In particular, Open AI make use of an autoregressive language model, which given a sequence of tokens (which can be thought of as pieces of text) tries to infer the next token, the inferred token is then added to the context and used to infer the next one and the process continues iteratively. In this way, an autoregressive model can be thought of as a model that uses previous predictions to generate new ones. This is useful as it has the capability to generate original mathematical terms, a capability which is lacking in conventional Automated Theorem Provers.</p><p>Open AI is specifically interested in this application as formal maths is “reasoning complete”, which refers to the fact that proving theorems requires general and flexible reasoning. Hence an advance in theorem proving is also an advance in reasoning more broadly. More importantly, automated theorem proving systems can quickly verify proofs which would aid the development of search methods. The ability to verify proofs also makes it possible to automatically generate new problems that could be used as training data which is particularly important considering that, outside of games like Chess and Go, it is difficult to find high-quality datasets for reasoning tasks.</p><p>Stanislas then went on to highlight that the large language models created by Open AI, such as GPT-3, are good at autocompletion and predicting output but are very bad at verifying their predictions, whereas formal systems are good at verifying predictions but are bad at generating new output. Naturally, this lead them to combine the two systems.</p><p><strong>Methodology &amp; Results</strong></p><p>The model was trained using the Metamath formal environment, a proof checker for verifying mathematical proofs. Metamath was chosen due to its large database of theorems that gives a straightforward way of extracting a dataset with compact subgoal representations which makes searching the proof tree straightforward. In simple terms, it yields a clean dataset of a goal with an accompanying proof step and a new subgoal produced by it. The dataset extracted from Metamath consisted of a Training set of 36k theorems with 3 million proof steps and a valid test set containing 1k theorems with 90k proof steps.</p><p>Stanislas further explains how pre-training was implemented in their model. Since the Metamath dataset was scarce they trained the model on a bigger dataset first and then re-trained it on the relevant (small) dataset and there was some effective transfer from one dataset to another.</p><p>As shown below, when the model was pre-trained on a dataset of 160 parameters from Common Crawl and then re-trained on Metamath there was a 4% increase(from 28.96% to 32.34%) in performance when compared to a model that is only trained on Metamath(trained from scratch).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fMIh5BXaJeiASjXZ.png" /><figcaption>Here, the sizes of the different models were compared. Generally, the bigger the model the better the performance and the lower the perplexity(a measure of the loss at the end of training).</figcaption></figure><p><strong>Learned Value Function</strong></p><p>Once the model is trained, the proofs that were closed (successfully completed) are examined and the goals that were explored are either labelled as useful or not. A new training objective is then introduced: to predict whether a goal is useful for the proof or not. Each goal is assigned a Negative or Positive label which is used to infer a learned value function that gives an estimate of how valuable a goal is to the proof.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/875/0*Cw6uZ4DbQAgFI8mw.png" /><figcaption>The learned value function allows the model to be trained on samples it generated and learn from its mistakes. With the learned value function, these results were achieved.</figcaption></figure><p>In each iteration, a proof is generated from the training set. The goals for the proof that was closed are then tagged whether they were useful or not (goals from proofs that were not closed are not tagged as they are all ‘negative’) and re-train the model from scratch on both the first objective and the second objective (proof step objective and the goal/outcome objective) and use the outcome objective to guide proof search which leads to a boost in performance. The process is iterated further for better performance as shown in the table above.</p><p><strong>Conclusion</strong></p><p>The previous SOTA (state of the art) for automated theorem proving known as MetaGen-IL reached a performance of 21.1% which Open AI managed to double using their transformer-based model approach. This suggests that transformer-based models are suitable for formal reasoning<strong>. </strong>Furthermore, when the trained model sampled proofs from all statements in Metamath it was able to generate shorter proofs for some of the statements which were then shared back with the Metamath community.</p><p>While we are still far from having a system capable of competing in Mathematics Olympiads, companies like Open AI continue to make strides towards building better models that give hope of one day creating an artificial mathematician.</p><p>A recording of a similar talk can be found on Youtube.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FzgGYyGkmWSw%3Fstart%3D2623%26feature%3Doembed%26start%3D2623&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DzgGYyGkmWSw&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FzgGYyGkmWSw%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7f696de5f6a8728338e898ff9bde3e6d/href">https://medium.com/media/7f696de5f6a8728338e898ff9bde3e6d/href</a></iframe><p>For more from Warwick AI, head to <a href="https://warwick.ai/">https://warwick.a</a>i.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cb589055685b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/warwick-artificial-intelligence/using-language-models-to-prove-truths-about-reality-cb589055685b">Using language models to prove truths about reality</a> was originally published in <a href="https://medium.com/warwick-artificial-intelligence">Warwick Artificial Intelligence</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>