When Truthful Representations Flip Under Deceptive Instructions?
"稀疏自编码器" 相关文章
Explaining GPT-2-Small Forward Passes with Edge-Level Autoencoder Circuits
From Messy Shelves to Master Librarians: Toy-Model Exploration of Block-Diagonal Geometry in LM Activations
L0 is not a neutral hyperparameter
Teach Old SAEs New Domain Tricks with Boosting
大模型知识回路的形成与SAE在可解释中的潜力丨周六直播·大模型可解释性读书会
Sparse Autoencoders for Sequential Recommendation Models: Interpretation and Flexible Control
大模型知识回路的形成与SAE在可解释中的潜力丨周四直播·大模型可解释性读书会
大模型知识回路的形成与SAE在可解释中的潜力丨周四直播·大模型可解释性读书会
Direct Preference Optimization Using Sparse Feature-Level Constraints
Feature Integration Spaces: Joint Training Reveals Dual Encoding in Neural Network Representations
苦研10年无果，千万经费打水漂，AI黑箱依然无解，谷歌撕破脸
苦研10年无果，千万经费打水漂！AI黑箱依然无解，谷歌撕破脸
苦研10年无果，千万经费打水漂！AI黑箱依然无解，谷歌撕破脸
Interpretable Fine Tuning Research Update and Working Prototype
Negative Results on Group SAEs
This AI Paper Introduces a Short KL+MSE Fine-Tuning Strategy: A Low-Cost Alternative to End-to-End Sparse Autoencoder Training for Interpretability
Takeaways From Our Recent Work on SAE Probing
Enhancing Instruction Tuning in LLMs: A Diversity-Aware Data Selection Strategy Using Sparse Autoencoders
Topological Data Analysis and Mechanistic Interpretability
