<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Vision-Language Models | Yiyang "Diana" Wang</title><link>https://hello-diana.github.io/tags/vision-language-models/</link><atom:link href="https://hello-diana.github.io/tags/vision-language-models/index.xml" rel="self" type="application/rss+xml"/><description>Vision-Language Models</description><generator>Hugo Blox Builder (https://hugoblox.com)</generator><language>en-us</language><lastBuildDate>Wed, 14 Oct 2026 00:00:00 +0000</lastBuildDate><image><url>https://hello-diana.github.io/media/icon_hu_9f6d321bccb06734.png</url><title>Vision-Language Models</title><link>https://hello-diana.github.io/tags/vision-language-models/</link></image><item><title>Do As I Do, Not As I Say: On-Policy Distillation Inherits the Teacher in Vision-Language Models</title><link>https://hello-diana.github.io/publication/on-policy-distillation/</link><pubDate>Wed, 14 Oct 2026 00:00:00 +0000</pubDate><guid>https://hello-diana.github.io/publication/on-policy-distillation/</guid><description>&lt;p&gt;This work compares &lt;strong&gt;on-policy distillation&lt;/strong&gt; with filtered supervised fine-tuning in vision-language models. It examines how teacher supervision on student-generated trajectories affects both answer accuracy and the ability to complete a response.&lt;/p&gt;
&lt;p&gt;The results show that on-policy supervision can transfer persistent teacher failure modes as well as knowledge, making teacher reliability on student-visited states central to successful distillation.&lt;/p&gt;</description></item></channel></rss>