adamkarvonen/activation_oracles

Official PyTorch Implementation for Learning a Generative Meta-Model of LLM Activations, ICML 2026

A library for mechanistic interpretability of GPT-style language models

Experiments on whether language models can learn to detect read-only activation probes

Extract residual-stream activations and apply steering vectors (including activation oracles) to any vLLM model during inference.

Keeping language models honest by directly eliciting knowledge encoded in their activations.

AllenAI's post-training codebase

Companion code for the global workspace interpretability paper

LLM training in simple, raw C/CUDA

Mapping out the "memory" of neural nets with data attribution