<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Latent-Space Reasoning | Yuwen Hao</title><link>http://page.haoyuwen.cn/tags/latent-space-reasoning/</link><atom:link href="http://page.haoyuwen.cn/tags/latent-space-reasoning/index.xml" rel="self" type="application/rss+xml"/><description>Latent-Space Reasoning</description><generator>HugoBlox Kit (https://hugoblox.com)</generator><language>en-us</language><lastBuildDate>Thu, 01 Jan 2026 00:00:00 +0000</lastBuildDate><image><url>http://page.haoyuwen.cn/media/icon_hu_da05098ef60dc2e7.png</url><title>Latent-Space Reasoning</title><link>http://page.haoyuwen.cn/tags/latent-space-reasoning/</link></image><item><title>SIRCL: Stabilizing Implicit Reasoning via Centripetal Latent Trajectories</title><link>http://page.haoyuwen.cn/publications/conference-paper/</link><pubDate>Thu, 01 Jan 2026 00:00:00 +0000</pubDate><guid>http://page.haoyuwen.cn/publications/conference-paper/</guid><description>
&lt;div class="callout flex px-4 py-3 mb-6 rounded-md border-l-4 bg-blue-100 dark:bg-blue-900 border-blue-500"
data-callout="note"
data-callout-metadata=""&gt;
&lt;span class="callout-icon pr-3 pt-1 text-blue-600 dark:text-blue-300"&gt;
&lt;svg height="24" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24"&gt;&lt;path fill="none" stroke="currentColor" stroke-linecap="round" stroke-linejoin="round" stroke-width="1.5" d="m16.862 4.487l1.687-1.688a1.875 1.875 0 1 1 2.652 2.652L6.832 19.82a4.5 4.5 0 0 1-1.897 1.13l-2.685.8l.8-2.685a4.5 4.5 0 0 1 1.13-1.897zm0 0L19.5 7.125"/&gt;&lt;/svg&gt;
&lt;/span&gt;
&lt;div class="callout-content dark:text-neutral-300"&gt;
&lt;div class="callout-title font-semibold mb-1"&gt;Note&lt;/div&gt;
&lt;div class="callout-body"&gt;&lt;p&gt;This paper is currently under review.&lt;/p&gt;&lt;/div&gt;
&lt;/div&gt;
&lt;/div&gt;
&lt;p&gt;SIRCL investigates instability in implicit chain-of-thought reasoning, where latent-token trajectories may drift away from the problem core as reasoning steps increase. The method introduces a sample-level center and hinge-style trust-region loss to regularize excessive latent displacement during training while keeping inference cost unchanged.&lt;/p&gt;</description></item></channel></rss>