MAxBench Framework Evaluates Multinomial Concept Recovery in LLMs
September 14, 2026
MAxBench is a geometry-agnostic evaluation framework designed to test how effectively models recover complex, non-binary concept representations. The study compares 10 localization methods across 6 concepts, finding that affine subspaces provide more reliable steering for multinomial categories than binary directions.
HOW THIS AFFECTS YOU
●
researcherYou can use this framework to move beyond binary refusal steering and evaluate more complex concept interventions.