• Understanding Language Models 2: Stable Features and Identifiable Causal Structure

  A stable feature basis is the prerequisite for interpretable causal structure: first demand run-to-run consistency, then recover temporal and instantaneous relations.

  11 min read   ·   April 02, 2026

  2026   ·   language-model   mechanistic-interpretability   causal-representation-learning   sparse-autoencoder   identifiability   ·   LLM   MI   CRL

• Understanding Language Models 1: Mechanistic Interpretability Meets Causal Representation Learning

  Mechanistic interpretability and causal representation learning study the same object, computation, but from complementary angles: circuits vs variables.

  9 min read   ·   February 26, 2026

  2026   ·   language-model   mechanistic-interpretability   causal-representation-learning   ·   LLM   MI   CRL

• Sequence Model 4: Nonstationary Dynamics

  Extending identifiability for sequence models to nonstationary dynamics.

  9 min read   ·   February 20, 2026

  2026   ·   sequence-model   identifiability   nonstationary   ·   CRL

• Sequence Model 3: Past Observations as Auxiliary Variables

  Deriving identifiability theorems for sequence models using the sufficient variability framework with past observations as auxiliary variables.

  5 min read   ·   February 04, 2026

  2026   ·   sequence-model   identifiability   ·   CRL

• Sequence Model 2: Sufficient Variability

  A family of assumptions that ensure identifiability in sequence models by leveraging sufficient variability in the latent dynamics.

  8 min read   ·   February 01, 2026

  2026   ·   sequence-model   identifiability   ·   CRL