← All demos
EMNLP 2026 Tutorial · Causal Foundations for Interpretability
Pearl's ladder sorts questions into rungs
Counterfactual: what this one input would have produced had
a
been different.
Set
a
to
ã
0.5
a
on 30 + 47
Rung 3 · Counterfactual
what this one input would have produced had
a
been different
Rung three adds the concept
a
stands for.
Rung 2 · Intervention
setting: does
y
change when we set
a
, over many inputs?
Rung 1 · Association
watching: does
y
go with
a
across inputs?
Eight sums, the observed activation a, the answer y, and the answer with a set to ã
rung-3 input
sum
carries?
a
observed
answer
y
y
with
a
:=
ã
a
: the activation at one layer and one position of a toy network that adds two numbers;
ã
: the value the slider sets it to.
The toy network adds a carry of 1 to the tens digit when
a
is above 0.5; the values of
a
are toy values.
The button on a row makes that sum the one input of rung 3.