MAxBench: A Multinomial Concept Recovery Benchmark
MAxBench is introduced as a geometry-agnostic evaluation framework for multinomial concept representations based on sampling from the recovered concept representation. It compares 10 localization methods covering 5 geometry types across 6 concepts and 4 models. Findings: affine subspaces steer more reliably and have greater recall than rank-one or linear subspaces; much of this advantage is due to better non-zero offsets rather than the choice of bases; manifold steering is competitive with the best methods when applicable.
MAxBench is a new benchmark for evaluating methods that recover multinomial concept representations in language models. It uses sampling from recovered representations to compare 10 localization methods across 5 geometry types, 6 concepts, and 4 models. Key findings indicate that affine subspaces outperform rank-one and linear subspaces in steering reliability and recall, largely due to better offsets, and manifold steering is competitive when applicable.
The benchmark's sampling-based evaluation is geometry-agnostic, allowing fair comparison across representation types. The advantage of affine subspaces over linear subspaces suggests that concept directions often require a non-zero offset, implying that concepts are not centered at the origin in activation space. Manifold methods, when applicable, can match the best affine methods, indicating that non-linear geometries may capture concept structure more accurately for some concepts.
Improved steering methods for multinomial concepts could enable more fine-grained control of language model behavior, which is valuable for applications requiring nuanced outputs (e.g., content moderation, style control, domain-specific generation). The finding that simple affine subspaces with learned offsets are effective suggests that practical steering tools can be implemented with relatively low complexity.
For AI companies, better concept steering can improve product safety and customization. The benchmark provides a standardized way to evaluate steering methods, potentially reducing R&D effort. The insight that offsets matter could lead to simpler, more robust steering implementations, lowering engineering costs.
Next signals to watch include: adoption of MAxBench in subsequent interpretability papers; extensions to more concepts and models; development of new localization methods that exploit offsets or manifold structures; and application of these steering techniques in production systems for controllable generation.