<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Mechanistic Interpretability — FoundationIndex</title>
    <link>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability</link>
    <description>A research discipline that reverse-engineers neural networks from the inside out—mapping hidden weights and activations into human-understandable concepts and circuits. Rather than judging an AI model only by its final outputs, mechanistic interpretability examines the internal algorithms it executes.</description>
    <language>en</language>
    <atom:link href="https://foundationindex.ustunozgur.com/s/mechanistic-interpretability/feed.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Mechanistic Interpretability · Paper · 2026-05-07</title>
      <link>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-nla-2026</link>
      <guid>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-nla-2026</guid>
      <pubDate>Thu, 07 May 2026 00:00:00 GMT</pubDate>
      <description>&lt;p&gt;Fraser-Taliente, Kantamneni, Ong, and colleagues introduced Natural Language Autoencoders (NLAs), jointly training an activation verbalizer and reconstructor with reinforcement learning so residual-stream states decode directly into natural-language explanations—surfacing unverbalized task strategies and evaluation awareness without manual feature labeling, alongside July 2026&amp;apos;s Jacobian-lens Global Workspace study (arXiv:2607.15495).&lt;/p&gt;&lt;p&gt;Sources: &lt;a href=&quot;https://transformer-circuits.pub/2026/nla/index.html&quot;&gt;Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations&lt;/a&gt;&lt;/p&gt;</description>
    </item>
    <item>
      <title>Mechanistic Interpretability · Release · 2025-12-19</title>
      <link>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-gemma-scope-2-2025</link>
      <guid>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-gemma-scope-2-2025</guid>
      <pubDate>Fri, 19 Dec 2025 00:00:00 GMT</pubDate>
      <description>&lt;p&gt;Google DeepMind released Gemma Scope 2 (alongside OpenAI&amp;apos;s November 2025 Weight-sparse transformers study, arXiv:2511.13653), open-sourcing JumpReLU sparse autoencoders, skip-transcoders, and cross-layer transcoders across Gemma 3 models up to 27B parameters so researchers can audit refusal, hallucination, and safety circuits.&lt;/p&gt;&lt;p&gt;Sources: &lt;a href=&quot;https://deepmind.google/blog/gemma-scope-2-helping-the-ai-safety-community-deepen-understanding-of-complex-language-model-behavior&quot;&gt;Gemma Scope 2: Helping the AI Safety Community Deepen Understanding of Complex Language Model Behavior&lt;/a&gt;&lt;/p&gt;</description>
    </item>
    <item>
      <title>Mechanistic Interpretability · Paper · 2025-03-27</title>
      <link>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-circuit-tracing-2025</link>
      <guid>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-circuit-tracing-2025</guid>
      <pubDate>Thu, 27 Mar 2025 00:00:00 GMT</pubDate>
      <description>&lt;p&gt;Ameisen, Lindsey, and Anthropic&amp;apos;s interpretability team published Circuit Tracing and On the Biology of a Large Language Model, using cross-layer transcoders (CLTs) to construct feature attribution graphs in Claude 3.5 Haiku—revealing multi-hop reasoning, advance rhyme planning, and shared cross-lingual concept spaces before open-sourcing circuit-tracer in May 2025.&lt;/p&gt;&lt;p&gt;Sources: &lt;a href=&quot;https://transformer-circuits.pub/2025/attribution-graphs/methods.html&quot;&gt;Circuit Tracing: Revealing Computational Graphs in Language Models&lt;/a&gt;, &lt;a href=&quot;https://transformer-circuits.pub/2025/attribution-graphs/biology.html&quot;&gt;On the Biology of a Large Language Model&lt;/a&gt;&lt;/p&gt;</description>
    </item>
    <item>
      <title>Mechanistic Interpretability · Paper · 2024-05-21</title>
      <link>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-scaling-monosemanticity-2024</link>
      <guid>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-scaling-monosemanticity-2024</guid>
      <pubDate>Tue, 21 May 2024 00:00:00 GMT</pubDate>
      <description>&lt;p&gt;Templeton et al. published Scaling Monosemanticity, applying dictionary learning to a frontier production model (Claude 3 Sonnet) to isolate millions of multimodal and multilingual features and verify causal steering by clamping individual directions—demonstrated publicly two days later as &amp;quot;Golden Gate Claude.&amp;quot;&lt;/p&gt;&lt;p&gt;Sources: &lt;a href=&quot;https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html&quot;&gt;Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet&lt;/a&gt;, &lt;a href=&quot;https://www.anthropic.com/news/golden-gate-claude&quot;&gt;Golden Gate Claude&lt;/a&gt;&lt;/p&gt;</description>
    </item>
    <item>
      <title>Mechanistic Interpretability · Paper · 2023-10-05</title>
      <link>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-monosemanticity-2023</link>
      <guid>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-monosemanticity-2023</guid>
      <pubDate>Thu, 05 Oct 2023 00:00:00 GMT</pubDate>
      <description>&lt;p&gt;Building on Toy Models of Superposition (arXiv:2209.10652), Bricken et al. published Towards Monosemanticity, using sparse autoencoders (SAEs) to unpack a polysemantic 512-neuron MLP layer into over 4,000 monosemantic dictionary features—isolating crisp directions for code syntax, DNA sequences, and specific scripts.&lt;/p&gt;&lt;p&gt;Sources: &lt;a href=&quot;https://transformer-circuits.pub/2023/monosemantic-features/index.html&quot;&gt;Towards Monosemanticity: Decomposing Language Models With Dictionary Learning&lt;/a&gt;, &lt;a href=&quot;https://www.anthropic.com/news/towards-monosemanticity-decomposing-language-models-with-dictionary-learning&quot;&gt;Towards Monosemanticity Announcement&lt;/a&gt;&lt;/p&gt;</description>
    </item>
    <item>
      <title>Mechanistic Interpretability · Paper · 2021-12-22</title>
      <link>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-transformer-circuits-2021</link>
      <guid>https://foundationindex.ustunozgur.com/s/mechanistic-interpretability#mech-interp-transformer-circuits-2021</guid>
      <pubDate>Wed, 22 Dec 2021 00:00:00 GMT</pubDate>
      <description>&lt;p&gt;Elhage, Olsson, Nanda, and colleagues at Anthropic published A Mathematical Framework for Transformer Circuits, modeling the Transformer residual stream as a linear communication bus and factoring attention heads into QK (what attends to what) and OV (what moves forward) matrices to explain induction heads in small models.&lt;/p&gt;&lt;p&gt;Sources: &lt;a href=&quot;https://transformer-circuits.pub/2021/framework/index.html&quot;&gt;A Mathematical Framework for Transformer Circuits&lt;/a&gt;&lt;/p&gt;</description>
    </item>
  </channel>
</rss>
