# Layer Sweep

## Metric
$$ completeness_{layer} = \frac{p_{layer}(object) - p_{-1}(object)}{p_{-1}(object)}$$

$$ contribution_{layer} = \frac{p_{layer}(object) - p_{layer - 1}(object)}{p_{-1}(object)}$$

restoring $h_{subject}$ at different layers in a corrupted run (different subject), but only at the `subject_last` position.
$$ causal\_score_{layer} = \frac{p(answer | h_{subject}^{layer} \text{ restored on corrupted run}) - p(answer | \text{clean run})}{p(answer | \text{clean run})}$$

| completeness | contribution | causal_score | faithfulness
| -------- | -------- | -------- | --------
|||||
||| |
||| |
||||
||||
||||
||||
||||

<!-- relation = `The capital of {} is`
Each calculated with 3 icl examples

$$ completeness_{layer} = \frac{p_{layer}(object) - p_{-1}(object)}{p_{-1}(object)}$$


---

$$ contribution_{layer} = \frac{p_{layer}(object) - p_{layer - 1}(object)}{p_{-1}(object)}$$


| J_norm | Jh_norm |
| -------- | -------- |
||
|

---

## Causal Tracing

## ICL-Mean estimation performance at different layers

-->