<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>weekly-papers-2026-10-24 on Sparse Notes</title>
    <link>https://sparsenotes.com/series/weekly-papers-2026-10-24/</link>
    <description>Recent content in weekly-papers-2026-10-24 on Sparse Notes</description>
    <image>
      <url>https://sparsenotes.com/images/og-default.png</url>
      <link>https://sparsenotes.com/images/og-default.png</link>
    </image>
    <generator>Hugo -- gohugo.io</generator>
    <lastBuildDate>Sat, 24 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://sparsenotes.com/series/weekly-papers-2026-10-24/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>A Fix for Softmax and Cross-Entropy</title>
      <link>https://sparsenotes.com/posts/2026/10/papers/softmax-cross-entropy-numerical-fix/</link>
      <pubDate>Sat, 24 Oct 2026 00:00:00 +0000</pubDate>
      
      <guid>https://sparsenotes.com/posts/2026/10/papers/softmax-cross-entropy-numerical-fix/</guid>
      <description>Barron, Gnanapragasam and Tschannen (Google DeepMind) find two floating-point bugs in every standard softmax/cross-entropy implementation. Fixing them in log space removes grokking&amp;#39;s Slingshot spikes, stops distilled students drifting from their teachers, and shows that logit margins under AdamW settle at log(1/ε).</description>
    </item>
    
    <item>
      <title>The End of Consensus Research?</title>
      <link>https://sparsenotes.com/posts/2026/10/papers/end-of-consensus-research-ai-greement/</link>
      <pubDate>Sat, 24 Oct 2026 00:00:00 +0000</pubDate>
      
      <guid>https://sparsenotes.com/posts/2026/10/papers/end-of-consensus-research-ai-greement/</guid>
      <description>Guerraoui, Kermarrec and colleagues at EPFL build a consensus-specific harness around a coding agent. In about 24 hours and $199 of tokens it produces A-greement, a geo-replicated protocol that cuts mean latency 12–33% against SwiftPaxos and Bodega, with a 98K-line Lean safety proof. Along the way it finds four unreported bugs in published protocols.</description>
    </item>
    
    <item>
      <title>The Working Set Model for Program Behavior (Denning, 1968)</title>
      <link>https://sparsenotes.com/posts/2026/10/papers/working-set-model-denning-1968/</link>
      <pubDate>Sat, 24 Oct 2026 00:00:00 +0000</pubDate>
      
      <guid>https://sparsenotes.com/posts/2026/10/papers/working-set-model-denning-1968/</guid>
      <description>Seminal paper of the week: Peter Denning defined a program&amp;#39;s working set as the pages it used in its last τ units of execution time, and argued the OS should only run programs whose working sets fit in memory. That rule is how operating systems stopped thrashing, and Linux and Windows still use the idea today.</description>
    </item>
    
    <item>
      <title>Weekly CS Paper Digest — 17 – 23 October 2026</title>
      <link>https://sparsenotes.com/posts/2026/10/papers/</link>
      <pubDate>Sat, 24 Oct 2026 00:00:00 +0000</pubDate>
      
      <guid>https://sparsenotes.com/posts/2026/10/papers/</guid>
      <description>This week&amp;#39;s picks: Google DeepMind&amp;#39;s fix for two floating-point bugs in every standard softmax and cross-entropy; EPFL&amp;#39;s agent harness that designed, built and proved a new geo-replicated consensus protocol; and Denning&amp;#39;s 1968 working set model as the seminal pick.</description>
    </item>
    
  </channel>
</rss>
