MAxBench: A Multinomial Concept Recovery Benchmark
This work introduces MAxBench, a geometry-agnostic evaluation framework for multinomial concept representations based on sampling from the recovered concept representation, and finds that affine subspaces steer more reliably and have greater recall than rank-one or linear subspaces.
Divya Appapogu, Freya Behrens, Yonatan Belinkov et al.
· 0 citations