Arrakis is a library to conduct, track and visualize mechanistic interpretability experiments.

[NeurIPS 2025 MechInterp Workshop - Spotlight] Official implementation of the paper "RelP: Faithful and Efficient Circuit Discovery in Language Models via Relevance Patching"

Lightweight representation engineering dataflow operations for agent developers.

Real-time 3D visualisation of SAE feature activations inside GPT-2, token by token

Investigating whether language models encode anticipated social consequences in their activations. Uses a 2x2 factorial design crossing truth × social valence to show that models are more sensitive to expected approval/disapproval than to truth itself.

Training and exploration of linear probes into Othello-GPT by Li et al. (2022)

Open-source EU AI Act Annex IV documentation toolkit. Mechanistic interpretability + circuit discovery for transformers. One function call generates a structured, hash-chained evidence package.

Implementation and analysis of Sparse Autoencoders for neural network interpretability research. Features interactive visualization dashboard and W&B integration.

Does Quantization Kill Interpretability? Scaling study across 5 models (124M-2.8B): RTN destroys induction heads in small models, GPTQ preserves them at all scales.

Evaluating how a model 'knowing what it knows' changes from base to instruct

Testing role-based pathways on small LLMs

Knowledge Activation Mapping & Understanding Interface (KAMUI) — A Transformer Interpretability Framework Built From Scratch in PyTorch.

A small, extensible mechanistic-interpretability lab — logit lens & activation patching on GPT-2 and Qwen3 behind a unified backend adapter. Config-driven, tested, laptop-friendly.

Mechanistic interpretability toolkit for comparing transformer activations, token shifts, and activation patching behaviour.

When does activation steering actually work? A reliability audit of steering vectors on GPT-2-small.

A Flax-based library for examining transformers, based on TransformerLens.

Reverse engineering the circuit responsible for the "greater than" capability in a language model

Probing where in Pythia's residual stream the decision to be sycophantic is already 'decided', using linear classifiers on per-layer activations against a small labeled sycophancy dataset.

Hands-on exploration of GPT-2 and transformer internals for text generation using TransformerLens — attention, mechanistic interpretability and sampling, explained step by step.

Logit Lens terminal visualizer (nostalgebraist, 2020) — decodes GPT-2's intermediate layer predictions using the unembedding matrix, built with TransformerLens and Rich.