A circuit-level map of a mid-size model

Anthropic's interpretability team traces how a model routes a multi-step arithmetic task — useful when debugging odd refusals.

The paper follows individual features through the network as the model adds two multi-digit numbers, showing separate paths for estimating the rough magnitude and computing the last digit.

The same tracing method is applied to cases where the model refuses a harmless request, locating the features that fire and the point at which the refusal decision is made.

The authors are clear that the maps cover a small fraction of the model's behaviour, but the tooling is released so others can apply it to their own questions.

A circuit-level map of a mid-size model

Key points