bottlenecks

Prediction Bottlenecks Fail to Uncover Causal Structure, New Benchmark Shows

A recent arXiv paper systematically tests the claim that prediction bottlenecks can discover causal structure. The claim does not survive: a linear bottleneck and tuned Lasso perform as well or better, and the reported intervention advantage is largely a sample-size confound. The benchmark itself emerges as the lasting artifact for causal testing in machine learning.