From Activation to Intervention: Interpreting Sparse Visual Features through Reconstruction
cs.CV
Submitted: 2026-11-03
Updated: 2026-11-03
Terminology
Sources
- Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering
- Toy Models of Superposition
- Explaining Classifiers with Causal Concept Effect (CaCE)
- Concept Steerers: Leveraging K-Sparse Autoencoders for Test-Time Controllable Generations
- Improved Baselines with Representation Autoencoders
- Interpretable and Testable Vision Features via Sparse Autoencoders
- SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features
- Revisiting the Importance of Individual Units in CNNs via Ablation
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs