hidden activations

3 papers

• Scaling can lead to compositional generalization
• The Emergence of Abstract Thought in Large Language Models Beyond Any Language
• VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set