<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
<channel>
  <title>MATRIX · DoubtLess</title>
  <link>https://derekwang2002.github.io/ai-intelligence/en/projects/doubtless/</link>
  <description>Material changes in AI models, agents, engineering and infrastructure, with primary sources and recommendations.</description>
  <language>en</language>
  <lastBuildDate>Mon, 02 Nov 2026 16:56:00 GMT</lastBuildDate>
  <atom:link href="https://derekwang2002.github.io/ai-intelligence/en/feeds/projects/doubtless.xml" rel="self" type="application/rss+xml"/>
  <item>
    <title>DoubtLess cuts reasoning tokens with contrastive decoding</title>
    <link>https://derekwang2002.github.io/ai-intelligence/en/events/ev-20261013-04/</link>
    <guid isPermaLink="false">ev-20261013-04-published</guid>
    <pubDate>Wed, 14 Oct 2026 15:57:53 GMT</pubDate>
    <category>New</category>
    <category>TRIAL</category>
    <description>&lt;p&gt;DoubtLess is a training-free decoding method that asks a small model to read the same chain of thought under brief and verbose instructions, then uses the logit difference to steer a larger model. The paper reports 31% fewer reasoning tokens on Qwen3-32B across seven benchmarks without lower aggregate accuracy, plus 20% fewer thinking tokens in coding self-debug tasks. Public code is available, but the results are author-reported and the reference setup requires an 80 GB GPU.&lt;/p&gt;&lt;p&gt;Recommendation: TRIAL&lt;/p&gt;&lt;p&gt;Sources: &lt;a href=&quot;https://arxiv.org/abs/2610.14745&quot;&gt;DoubtLess: Efficient Reasoning via Contrastive Decoding&lt;/a&gt; · &lt;a href=&quot;https://github.com/zhangshaowei1023/DoubtLess&quot;&gt;DoubtLess repository&lt;/a&gt;&lt;/p&gt;</description>
  </item>
</channel>
</rss>
