Definition
Plain language
A method that figures out which connections inside a transformer matter for a behavior, with just two forward-pass-style runs.
As stated in the literature
Position-aware Edge Attribution Patching — a gradient-based circuit-tracing technique that estimates the causal importance of every edge in a model's computation graph from a clean-versus-corrupted pair.
Also called: Position-aware Edge Attribution Patching
Why it matters: Cheap circuit attribution makes it feasible to map causal structure inside large models without prohibitive per-edge intervention costs.
For example, PEAP scores every edge in the model's computation graph for how much it matters to getting the correct answer, using just a clean run and a corrupted one.
Heard on the show
“The method they use is called Position-aware Edge Attribution Patching — PEAP, if you like acronyms — and the intuition is pretty clean.”Episode 055 — Why LLM Judges Flip Their Verdicts When You Change the Question Format