MAxBench: A Multinomial Concept Recovery Benchmark
MAxBench evaluates multinomial concept recovery methods, finding affine subspaces steer most reliably but none consistently beats prompting.
MAxBench is a geometry-agnostic evaluation framework for multinomial concept representations in language models, based on sampling from recovered concept representations. It compares 10 localization methods covering 5 geometry types across 6 concepts and 4 models. Findings show affine subspaces steer more reliably than rank-one or linear subspaces due to better non-zero offsets, manifold steering is competitive where applicable, and no method consistently outperforms prompting.
- Compares 10 localization methods across 5 geometry types, 6 concepts, and 4 models
- Affine subspaces steer more reliably with better non-zero offsets
- No method consistently outperforms prompting for concept steering
From the original article
1 min read at arxiv.orgFine-grained control of language model behaviors (e. g. , steering) is among the more actionable outcomes of interpretability research.
By Divya Appapogu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.