CircuitLasso preprint reports lower computing costs for mapping AI model behavior

Researchers report that CircuitLasso, a method for mapping how language-model components produce predictions, matches the structural accuracy of leading intervention-based methods on benchmark data at a fraction of their computational cost. The preprint uses sparse linear regression to learn circuits linking features extracted by sparse autoencoders, which help distinguish meanings that individual neurons can mix together. The researchers also used insights from those circuits to achieve comparable performance at substantially lower cost on a task testing generalization across domains. The reported accuracy comparison is limited to the benchmark data.