Reproduction Gallery
End-to-end reproductions of published interpretability results, implemented with Murano pipelines. Each entry includes a runnable notebook, paper reference, and walkthrough.
Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 Small
Reverse-engineers how GPT-2 small identifies the indirect object, localizing the name-mover, negative name-mover, and S-inhibition heads. Reproduced with Murano path patching, attention analysis, and OV circuits.
The Geometry of Truth: Emergent Linear Structure in LLM Representations of True/False Datasets
Shows that the truth of a factual statement is encoded as a single linear direction in the residual stream: true and false statements separate, a mass-mean probe reads the direction, it generalizes across topics, and adding it flips the model's judgement. Reproduced with Murano recording, probing, and interventions.
Function Vectors in Large Language Models
Shows that a small set of attention heads carry a compact, portable representation of an in-context task: summing their outputs gives a function vector, and adding it to a zero-shot prompt makes GPT-J perform the task. Reproduced with Murano recording, causal mediation, and interventions.
Shape Happens: Automatic Feature Manifold Discovery in LLMs via Supervised Multi-Dimensional Scaling
Introduces SMDS, a framework for discovering and comparing geometric hypotheses about how LLMs organize features in latent space. Shows that feature manifolds form circles, chains, and clusters reflecting the semantic structure of the encoded concept.