Steering-first neuron explorer for Qwen3.5-4B: push every MLP neuron, keep every rollout, browse the results
Mechanistic Interpretability GitHub Repositories (2026) · Open Source Tools
Discover the best Mechanistic Interpretability open source repositories on GitHub. Explore trending tools, live star counts, source code, and developer projects updated daily.
Run, serve, query and edit AI models — a Rust model runtime built on VINDEX3, an executable, inspectable model format.
Laplace's Scissors: protected readout rotations for open LLMs. Diagnose weak output directions, repair answers, preserve successful generations, and analyze scaling deviations with Rust and Triton on CUDA/HIP.
Dyno Lab — an AI safety, alignment and interpretability research workbench for Apple Silicon. Explore activations, probes, SAEs and interventions with local inference, a Python SDK, APIs and MCP.