<?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>Interpretability on TPOW Lab</title><link>https://tpow-001.netlify.app/tags/interpretability/</link><description>Recent content in Interpretability on TPOW Lab</description><generator>Hugo</generator><language>en</language><copyright>Copyright &amp;copy; 2025-2026 TPOW-001. All Rights Reserved.</copyright><lastBuildDate>Thu, 28 May 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://tpow-001.netlify.app/tags/interpretability/index.xml" rel="self" type="application/rss+xml"/><item><title>llm-neuron-atlas 詳細教學 — 3D LLM Neuron Atlas Tutorial</title><link>https://tpow-001.netlify.app/post/2026-05-28-llm-neuron-atlas-tutorial/</link><pubDate>Thu, 28 May 2026 00:00:00 +0000</pubDate><guid>https://tpow-001.netlify.app/post/2026-05-28-llm-neuron-atlas-tutorial/</guid><description><![CDATA[<h1 id="llm-neuron-atlas-詳細教學" data-numberify>llm-neuron-atlas 詳細教學<a class="anchor ms-1" href="#llm-neuron-atlas-詳細教學"></a></h1>
<blockquote>
<p>Live demo: <a href="https://charenix.com/qwen3b-atlas" target="_blank" rel="noopener noreferrer">https://charenix.com/qwen3b-atlas<i class="fas fa-external-link-square-alt ms-1"></i></a>
Repo: <a href="https://github.com/norika1207-lab/llm-neuron-atlas" target="_blank" rel="noopener noreferrer">https://github.com/norika1207-lab/llm-neuron-atlas<i class="fas fa-external-link-square-alt ms-1"></i></a>
作者: Norika Oda (ORCID 0009-0006-6816-9891)
License: MIT</p></blockquote>

<h2 id="1-專案定位" data-numberify>1. 專案定位<a class="anchor ms-1" href="#1-專案定位"></a></h2>
<p><code>llm-neuron-atlas</code> 屬於 <strong>mechanistic interpretability (機械式可解釋性)</strong> 工具範疇，但與既有工具有結構性的差異：</p>
<table>
  <thead>
      <tr>
          <th>工具</th>
          <th>視角</th>
          <th>規模</th>
          <th>互動</th>
      </tr>
  </thead>
  <tbody>
      <tr>
          <td>BertViz</td>
          <td>attention 矩陣</td>
          <td>單 head / 單層</td>
          <td>2D</td>
      </tr>
      <tr>
          <td>Neuronpedia</td>
          <td>SAE feature dashboard</td>
          <td>單 feature</td>
          <td>2D 圖卡</td>
      </tr>
      <tr>
          <td>Anthropic circuits</td>
          <td>手繪電路圖</td>
          <td>局部 (~10 nodes)</td>
          <td>靜態</td>
      </tr>
      <tr>
          <td><strong>llm-neuron-atlas</strong></td>
          <td><strong>per-neuron + 全層 + cross-arch</strong></td>
          <td><strong>73,728 nodes / 716K edges</strong></td>
          <td><strong>3D 即時</strong></td>
      </tr>
  </tbody>
</table>
<p><strong>核心定位</strong>：把整個 transformer 的「神經元城市」(neuron-as-city / weight-as-road) 一次性渲染出來，讓使用者用空間導航的方式探索 LLM 的內部結構。</p>]]></description></item></channel></rss>