Explaining How Visual, Textual and Multimodal Encoders Share Concepts
"稀疏自动编码器" 相关文章
On the transferability of Sparse Autoencoders for interpreting compressed models
Causal Language Control in Multilingual Transformers via Sparse Feature Steering
Are SAE features from the Base Model still meaningful to LLaVA?
解释器模型首创！Tilde打破提示工程局限，让AI推理更精准
解释器模型首创！Tilde打破提示工程局限，让AI推理更精准
Google DeepMind has a new way to look inside an AI’s “mind”
Peering Inside AI: How DeepMind’s Gemma Scope Unlocks the Mysteries of AI
SPARE: Training-Free Representation Engineering for Managing Knowledge Conflicts in Large Language Models
EIS XIV: Is mechanistic interpretability about to be practically useful?
Domain-specific SAEs
An X-Ray is Worth 15 Features: Sparse Autoencoders for Interpretable Radiology Report Generation
An X-Ray is Worth 15 Features: Spare Autoencoders for Interpretable Radiology Report Generation
AI Safety at the Frontier: Paper Highlights, August '24
Evaluating Sparse Autoencoders with Board Game Models
Google Deepmind Researchers Introduce Jumprelu Sparse Autoencoders: Achieving State-of-the-Art Reconstruction Fidelity
Initial Experiments Using SAEs to Help Detect AI Generated Text
BatchTopK: A Simple Improvement for TopK-SAEs
SAEs (usually) Transfer Between Base and Chat Models
Deceptive agents can collude to hide dangerous features in SAEs
