<?xml version="1.0" encoding="UTF-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>Junjie Xu — Work Blog</title>
  <subtitle>Notes on project progress, research practice, and human-AI collaboration.</subtitle>
  <id>https://andrejjxu.github.io/writing/</id>
  <updated>2026-08-01T00:00:00Z</updated>
  <link rel="alternate" type="text/html" hreflang="en" href="https://andrejjxu.github.io/writing/" />
  <link rel="self" type="application/atom+xml" href="https://andrejjxu.github.io/feed.xml" />
  <author><name>Junjie Xu</name></author>
  <entry>
    <title>Alignment is more than making a model understand music</title>
    <id>https://andrejjxu.github.io/writing/aesthetic-alignment/</id>
    <updated>2026-08-01T00:00:00Z</updated>
    <link rel="alternate" type="text/html" href="https://andrejjxu.github.io/writing/aesthetic-alignment/" />
    <summary>Music-driven image generation is not only a cross-modal mapping problem. It is also about turning rhythm, tension, and affective texture into stable visual decisions that people can perceive.</summary>
  </entry>
  <entry>
    <title>From affective intent to visual decisions: Notes on music-to-image generation</title>
    <id>https://andrejjxu.github.io/writing/music-to-image-project-log/</id>
    <updated>2026-08-01T00:00:00Z</updated>
    <link rel="alternate" type="text/html" href="https://andrejjxu.github.io/writing/music-to-image-project-log/" />
    <summary>Affective music-to-image generation is not a matter of replacing music labels with image labels. The real question is how rhythm, timbre, and human intention can jointly shape composition, style, and emotional intensity.</summary>
  </entry>
  <entry>
    <title>Why comfort deserves a place in generative model evaluation</title>
    <id>https://andrejjxu.github.io/writing/comfort-benchmark/</id>
    <updated>2026-07-01T00:00:00Z</updated>
    <link rel="alternate" type="text/html" href="https://andrejjxu.github.io/writing/comfort-benchmark/" />
    <summary>When an image is used for companionship, relief, or intervention, realism and beauty are not enough. We also need to ask whether it leaves a person room to breathe.</summary>
  </entry>
  <entry>
    <title>In digital therapeutics, define what support means first</title>
    <id>https://andrejjxu.github.io/writing/digital-therapeutics-project-log/</id>
    <updated>2026-07-01T00:00:00Z</updated>
    <link rel="alternate" type="text/html" href="https://andrejjxu.github.io/writing/digital-therapeutics-project-log/" />
    <summary>Multimodal digital therapeutics should not optimize for generation quality alone. A system becomes supportive only when effectiveness is reconsidered through context, emotional load, and human autonomy.</summary>
  </entry>
  <entry>
    <title>Human feedback should not be the final column in a training set</title>
    <id>https://andrejjxu.github.io/writing/human-in-loop/</id>
    <updated>2026-06-01T00:00:00Z</updated>
    <link rel="alternate" type="text/html" href="https://andrejjxu.github.io/writing/human-in-loop/" />
    <summary>In human-AI systems, people are not only annotators who score model outputs. They also define what is effective, appropriate, and worth exploring next.</summary>
  </entry>
  <entry>
    <title>Working on EduChat-R1: How open collaboration reaches real settings</title>
    <id>https://andrejjxu.github.io/writing/educhat-r1-open-source-log/</id>
    <updated>2026-06-01T00:00:00Z</updated>
    <link rel="alternate" type="text/html" href="https://andrejjxu.github.io/writing/educhat-r1-open-source-log/" />
    <summary>EduChat-R1 clarified how open research works for me: model capability, engineering collaboration, and the needs of real users must be calibrated in the same feedback loop.</summary>
  </entry>
  <entry>
    <title>Leave research room to slow down</title>
    <id>https://andrejjxu.github.io/writing/research-rhythm/</id>
    <updated>2026-05-01T00:00:00Z</updated>
    <link rel="alternate" type="text/html" href="https://andrejjxu.github.io/writing/research-rhythm/" />
    <summary>In research shaped by metrics and deadlines, slowing down is not a loss of efficiency. It gives the real questions room to appear.</summary>
  </entry>
</feed>
