IWIBench is a new benchmark for measuring how well an editor model can modify the weights of a target model, such as Gemma 3 4B IT, without using gradients. It tests the ability to alter lexical associations or factual knowledge while preserving original model capabilities.
HOW THIS AFFECTS YOU
●
builderThis opens up ways to update model knowledge or formats by editing weights directly rather than retraining.
●
researcherYou can use this to test how well one model can edit the weights of another.