Imagen de Portada

Tokenized SAEs: Disentangling SAE Reconstructions

Fuente: arXiv

Guardado en:

Detalles Bibliográficos
Autores principales:	Dooms, Thomas, Wilhelm, Daniel
Formato:	Preprint
Publicado:	2025
Materias:	Machine Learning
Acceso en línea:	Acceder al recurso
Etiquetas:	Agregar Etiqueta Sin Etiquetas, Sea el primero en etiquetar este registro!

Ejemplares similares

Interpretability as Compression: Reconsidering SAE Explanations of Neural Activations with MDL-SAEs
por: Ayonrinde, Kola, et al.
Publicado: (2024)

Finding Manifolds With Bilinear Autoencoders
por: Dooms, Thomas, et al.
Publicado: (2025)

Distribution-Aware Feature Selection for SAEs
por: Oozeer, Narmeen, et al.
Publicado: (2025)

The Rate-Distortion-Polysemanticity Tradeoff in SAEs
por: Mencattini, Tommaso, et al.
Publicado: (2026)

Analyzing (In)Abilities of SAEs via Formal Languages
por: Menon, Abhinav, et al.
Publicado: (2024)

Fractals made Practical: Denoising Diffusion as Partitioned Iterated Function Systems
por: Dooms, Ann
Publicado: (2026)

Weight-based Decomposition: A Case for Bilinear MLPs
por: Pearce, Michael T., et al.
Publicado: (2024)

Investigating Sensitive Directions in GPT-2: An Improved Baseline and Comparative Analysis of SAEs
por: Lee, Daniel J., et al.
Publicado: (2024)

Bilinear autoencoders find interpretable manifolds
por: Dooms, Thomas, et al.
Publicado: (2026)

Residual Stream Analysis with Multi-Layer SAEs
por: Lawson, Tim, et al.
Publicado: (2024)

Compositionality Unlocks Deep Interpretable Models
por: Dooms, Thomas, et al.
Publicado: (2025)

Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines?
por: Korznikov, Anton, et al.
Publicado: (2026)

Ablating Archetypes: The Stability of Archetypal SAEs is an Artifact of Initialization and Metric Design
por: Brzozowski, Michał, et al.
Publicado: (2026)

Resa: Transparent Reasoning Models via SAEs
por: Wang, Shangshang, et al.
Publicado: (2025)

Can SAEs reveal and mitigate racial biases of LLMs in healthcare?
por: Ahsan, Hiba, et al.
Publicado: (2025)

Features Emerge as Discrete States: The First Application of SAEs to 3D Representations
por: Miao, Albert, et al.
Publicado: (2025)

SAEs Are Good for Steering -- If You Select the Right Features
por: Arad, Dana, et al.
Publicado: (2025)

Teach Old SAEs New Domain Tricks with Boosting
por: Koriagin, Nikita, et al.
Publicado: (2025)

From Mechanistic to Compositional Interpretability
por: Gauderis, Ward, et al.
Publicado: (2026)

Bilinear MLPs enable weight-based mechanistic interpretability
por: Pearce, Michael T., et al.
Publicado: (2024)

Position: Mechanistic Interpretability Should Prioritize Feature Consistency in SAEs
por: Song, Xiangchen, et al.
Publicado: (2025)

Evolution of SAE Features Across Layers in LLMs
por: Balcells, Daniel, et al.
Publicado: (2024)

Mechanistic Interpretability with SAEs: Probing Religion, Violence, and Geography in Large Language Models
por: Simbeck, Katharina, et al.
Publicado: (2025)

Evaluating SAE interpretability without explanations
por: Paulo, Gonçalo, et al.
Publicado: (2025)

SAE: Single Architecture Ensemble Neural Networks
por: Ferianc, Martin, et al.
Publicado: (2024)

Behavior Tokens Speak Louder: Disentangled Explainable Recommendation with Behavior Vocabulary
por: Feng, Xinshun, et al.
Publicado: (2025)

When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability
por: Gonzalez, ML Nissen, et al.
Publicado: (2026)

Deep Reinforcement Learning for Local Path Following of an Autonomous Formula SAE Vehicle
por: Merton, Harvey, et al.
Publicado: (2024)

SCALAR: Benchmarking SAE Interaction Sparsity in Toy LLMs
por: Fillingham, Sean P., et al.
Publicado: (2025)

MetaSAEs: Joint Training with a Decomposability Penalty Produces More Atomic Sparse Autoencoder Latents
por: Levinson, Matthew
Publicado: (2026)

OrtSAE: Orthogonal Sparse Autoencoders Uncover Atomic Features
por: Korznikov, Anton, et al.
Publicado: (2025)

Concept-SAE: Active Causal Probing of Visual Model Behavior
por: Ding, Jianrong, et al.
Publicado: (2025)

Evaluating Synthetic Activations composed of SAE Latents in GPT-2
por: Giglemiani, Giorgi, et al.
Publicado: (2024)

Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders
por: Cao, Tue M., et al.
Publicado: (2026)

AlignSAE: Concept-Aligned Sparse Autoencoders
por: Yang, Minglai, et al.
Publicado: (2025)

SAEs $\textit{Can}$ Improve Unlearning: Dynamic Sparse Autoencoder Guardrails for Precision Unlearning in LLMs
por: Muhamed, Aashiq, et al.
Publicado: (2025)

Dense SAE Latents Are Features, Not Bugs
por: Sun, Xiaoqing, et al.
Publicado: (2025)

HH-SAE: Discovering and Steering Hierarchical Knowledge of Complex Manifolds
por: Wu, Honghan, et al.
Publicado: (2026)

TimeSAE: Sparse Decoding for Faithful Explanations of Black-Box Time Series Models
por: Oublal, Khalid, et al.
Publicado: (2026)

SAE-FD: Sparse Autoencoder Feature Distillation for Continual Learning of Large Language Models
por: Zhang, Mingxu, et al.
Publicado: (2026)