SPARC: Concept-Aligned Sparse Autoencoders for Cross-Model and Cross-Modal Interpretability

Researchers introduced SPARC, a framework that creates unified latent spaces across different AI models and modalities, enabling direct comparison of how various architectures represent identical concepts. The method achieves 0.80 Jaccard similarity on Open Images, tripling alignment compared to previous methods, and enables practical applications like text-guided spatial localization in vision-only models.

• →SPARC creates a single unified latent space shared across diverse AI architectures and modalities like vision and multimodal models.
• →The framework uses Global TopK sparsity and Cross-Reconstruction Loss to ensure semantic consistency between different models.
• →SPARC achieves 0.80 Jaccard similarity on Open Images, more than tripling concept alignment compared to existing methods.
• →The technology enables practical applications including text-guided spatial localization and cross-model retrieval.
• →Individual dimensions in SPARC's shared space correspond to similar high-level concepts across different models and modalities.