Audio-Visual-Language Learning · Interactive AI · Mechanistic Interpretability
A sparse-autoencoder training setup that aligns dictionary directions with vocabulary anchors, then checks named features through token examples and reconstruction behavior.
A survey and resource map that organizes AVL work by task setup, modality alignment, benchmark coverage, evaluation metrics, and gaps in current datasets.
A controlled Blocks World benchmark that separates object-presence cues from relational prompts, testing whether CLIP distinguishes spatial relations in synthetic scenes.
