Does composing function vectors compose their behaviors? Eight-stage causal pipeline with CIE head ranking, layer sweeps, and multi-position patching.
-
Updated
Aug 17, 2026 - Python
#
function-vectors
Here are 2 public repositories matching this topic...
Localising and editing LLM traits and fine-tuned habits via component-level causal patching. Demonstrates repairing misaligned LoRAs by swapping 17 attention heads with clean base model activations.
alignment mechanistic-interpretability qwen llm-safety llama3 causal-patching function-vectors model-diff
-
Updated
Aug 2, 2026 - Jupyter Notebook
Add this topic to your repo
To associate your repository with the function-vectors topic, visit your repo's landing page and select "manage topics."