LLM Hub — The Transformer Circuits Thread

The Transformer Circuits Thread

Anthropic's mechanistic interpretability research, for engineers. Six deep-dive decks trace one continuous arc — from the linear-algebra framework that made attention readable, through superposition and the dictionary-learning breakthrough that resolves it, to tracing the step-by-step computation of a deployed Claude model. Each takes one landmark publication and answers what it solves, how the mechanism works, and why it matters to building and operating LLM/agent systems.

Residual StreamQK / OV CircuitsInduction HeadsSuperpositionSparse AutoencodersAttribution Graphs6 publications

Presentations in This Series

  1. A Mathematical Framework for Transformer Circuits →
    The residual stream as a communication channel; attention heads split into independent QK (where to attend) and OV (what to write) circuits; path expansion and the model's hidden linear structure; how composition in two-layer models produces induction heads.
    liveElhage et al. — Dec 2021
  2. In-context Learning & Induction Heads →
    The prefix-match-then-copy mechanism; the abrupt phase change in training where induction heads form; the in-context-learning score; and the multiple lines of evidence that induction heads are the primary engine of in-context learning.
    liveOlsson et al. — Mar 2022
  3. Toy Models of Superposition →
    Features as directions; how a network packs more features than it has neurons by exploiting sparsity and tolerating interference; the phase transitions and clean geometry of superposition; and why this makes neurons polysemantic and interpretability hard.
    liveElhage et al. — Sep 2022
  4. Towards Monosemanticity →
    Sparse autoencoders / dictionary learning decompose a one-layer model into thousands of interpretable, causal, monosemantic features where the raw neurons were not; specificity and causal tests; feature splitting and cross-model universality.
    liveBricken et al. — Oct 2023
  5. Scaling Monosemanticity →
    The recipe scaled to the production model Claude 3 Sonnet: millions of abstract, multilingual, multimodal features; safety-relevant features (deception, sycophancy, code backdoors); causal feature steering and the famous "Golden Gate Claude".
    liveTempleton et al. — May 2024
  6. On the Biology of a Large Language Model →
    Circuit tracing with cross-layer transcoders and attribution graphs on Claude 3.5 Haiku: multi-step reasoning, planning ahead in poems, shared multilingual circuits, mental arithmetic, hallucination, refusals and unfaithful chain-of-thought — a mechanistic microscope on a deployed model.
    liveLindsey, Ameisen et al. — Mar 2025

How to read this series. The six decks are chronological and build on one another — the framework gives the vocabulary, induction heads are the first real circuit, superposition explains why interpretation is hard, dictionary learning and its scaling resolve superposition into readable features, and circuit tracing uses those features to watch a deployed model think. Each deck also stands alone, with a diagram of the key idea wherever a picture helps.