Definition
Plain language
A benchmark that gives an AI a description of statistical correlations and asks it to deduce what causes what.
As stated in the literature
A causal-inference benchmark presenting correlational and conditional-independence facts in text and asking the model to judge causal claims; its extended variant scales to 24 variables.
Also called: Extended Corr2Cause
Why it matters: It probes whether a model can move beyond noticing correlations to genuinely deducing cause and effect from described patterns.
For example, it tells a model that two things tend to rise together along with a third, then asks it to reason out which one actually drives the others.
Heard on the show
“So the benchmark that started this line of work is called Corr2Cause.”Episode 091 — When Better Fine-Tuning Can't Help: A Geometric Impossibility in LLM Causal Reasoning