WriteSAE introduces sparse autoencoders with rank-1 matrix atoms for recurrent state updates, allowing replacement tests that outperform deletion on 92.4% of positions and a formula predicting logit changes with R²=0.98.
hub Canonical reference
Improving Dictionary Learning with Gated Sparse Autoencoders
Canonical reference. 83% of citing Pith papers cite this work as background.
23
Pith papers citing it
Background
83% of classified citations
abstract
Recent work has found that sparse autoencoders (SAEs) are an effective technique for unsupervised discovery of interpretable features in language models' (LMs) activations, by finding sparse, linear reconstructions of LM activations. We introduce the Gated Sparse Autoencoder (Gated SAE), which achieves a Pareto improvement over training with prevailing methods. In SAEs, the L1 penalty used to encourage sparsity introduces many undesirable biases, such as shrinkage -- systematic underestimation of feature activations. The key insight of Gated SAEs is to separate the functionality of (a) determining which directions to use and (b) estimating the magnitudes of those directions: this enables us to apply the L1 penalty only to the former, limiting the scope of undesirable side effects. Through training SAEs on LMs of up to 7B parameters we find that, in typical hyper-parameter ranges, Gated SAEs solve shrinkage, are similarly interpretable, and require half as many firing features to achieve comparable reconstruction fidelity.
hub tools
citation-role summary
background 5
method 1
citation-polarity summary
representative citing papers
SASA replaces single-vector decoders in SAEs with learned subspaces plus block sparsity and nuclear-norm regularization, proving that a single group becomes the global minimizer once block size meets intrinsic dimension and yielding polynomial rather than exponential sample complexity.
SA-GSAE with Bi-Jump-ReLU enables one latent to encode both polarities of anticorrelated features, Pareto-dominating or matching full-width gated SAEs while reducing dead latents by up to 500x on some LLM hookpoints.
CEDAR learns an invertible rotation of vision-language embeddings to concentrate semantics into sparse, axis-aligned coordinates for improved interpretability.
SLIM decomposes LLM hidden states via sparse autoencoders with learnable gates to enable precise, interpretable steering of molecular properties, yielding up to 42.4-point gains on the MolEditRL benchmark.
HH-SAE factorizes manifolds into nested contextual (L0), atomic (f1), and compository (f2) tiers, achieving 0.9156 cross-domain zero-shot AUC in fraud detection and +9.9% AUPRC lift in steered synthesis.
A cross-attention SAE with sparsemax attention achieves lower reconstruction loss and higher-quality concepts than fixed-sparsity baselines by making activation counts data-dependent.
Joint training of a primary SAE with a meta SAE that applies a decomposability penalty on decoder directions produces more atomic latents, shown by 7.5% lower mean absolute phi and 7.6% higher fuzzing scores on GPT-2.
K-sparse autoencoders with dead-latent fixes produce clean scaling laws and better feature quality metrics that improve with size, shown by training a 16-million-latent model on GPT-4 activations.
Sparse autoencoders scaled to 34 million features on Claude 3 Sonnet yield interpretable, steerable representations of concrete and abstract concepts that generalize across languages and modalities.
DECO is a sparse MoE architecture with ReLU-based routing, learnable expert scaling, and NormSiLU activation that matches dense Transformer performance at 20% expert activation and delivers 2.93x speedup on Jetson AGX Orin.
Health foundation model embeddings contain an interpretable symbolic organization shared across modalities that supports cross-domain transfer without joint training.
Sparse autoencoders isolate unstable features in reward model representations and enable two mitigation techniques that reduce preference errors on perturbed inputs without retraining.
Graph-motif clustering of SAE features via a frequency-binned WL kernel recovers structural families not captured by decoder cosine similarity or token histograms.
Adaptive elastic net SAEs (AEN-SAEs) mitigate feature starvation in SAEs by combining ℓ2 structural stability with adaptive ℓ1 reweighting, producing a Lipschitz-continuous sparse coding map that recovers global feature support under mild assumptions.
DACO curates a 15,000-concept dictionary from 400K image-caption pairs and uses it to initialize an SAE that enables granular, concept-specific steering of MLLM activations, raising safety scores on MM-SafetyBench and JailBreakV while preserving general capabilities.
Sparse autoencoders enable phase synchronization in frozen graph CFD surrogates through Hilbert-identified oscillatory features and SVD-based time-varying rotations.
VPiT enables pretrained LLMs to perform both visual understanding and generation by predicting discrete text tokens and continuous visual tokens, with understanding data proving more effective than generation-specific data.
Sparse feature circuits are introduced as interpretable causal subnetworks in language models, supporting unsupervised discovery of thousands of circuits and a method called SHIFT to improve classifier generalization by ablating irrelevant features.
Prototype-Based Sparse Steering decomposes query activations with SAEs and optimizes sparse features via gradients to steer LLM outputs toward specific behaviors.
A five-stage causal feature analysis methodology is proposed and tested on GPT-2 for IOI, showing partial causality of SAE features, robustness differences under shifts, and deployment cost benefits.
RET learns temporally consistent macrovariables from LLM activations via self-supervised learning to support interpretability, early behavioral prediction, and causal intervention.
The survey organizes mechanistic interpretability techniques into a Locate-Steer-Improve framework to enable actionable improvements in LLM alignment, capability, and efficiency.
citing papers explorer
-
WriteSAE: Sparse Autoencoders for Recurrent State
WriteSAE introduces sparse autoencoders with rank-1 matrix atoms for recurrent state updates, allowing replacement tests that outperform deletion on 92.4% of positions and a formula predicting logit changes with R²=0.98.
-
Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability
SASA replaces single-vector decoders in SAEs with learned subspaces plus block sparsity and nuclear-norm regularization, proving that a single group becomes the global minimizer once block size meets intrinsic dimension and yielding polynomial rather than exponential sample complexity.
-
Sign-Aware Gated Sparse Autoencoders: Modeling Anticorrelated Features with Bi-Jump-ReLU Activations
SA-GSAE with Bi-Jump-ReLU enables one latent to encode both polarities of anticorrelated features, Pareto-dominating or matching full-width gated SAEs while reducing dead latents by up to 500x on some LLM hookpoints.
-
Conceptualizing Embeddings: Sparse Disentanglement for Vision-Language Models
CEDAR learns an invertible rotation of vision-language embeddings to concentrate semantics into sparse, axis-aligned coordinates for improved interpretability.
-
SLIM: Sparse Latent Steering for Interpretable and Property-Directed LLM-Based Molecular Editing
SLIM decomposes LLM hidden states via sparse autoencoders with learnable gates to enable precise, interpretable steering of molecular properties, yielding up to 42.4-point gains on the MolEditRL benchmark.
-
HH-SAE: Discovering and Steering Hierarchical Knowledge of Complex Manifolds
HH-SAE factorizes manifolds into nested contextual (L0), atomic (f1), and compository (f2) tiers, achieving 0.9156 cross-domain zero-shot AUC in fraud detection and +9.9% AUPRC lift in steered synthesis.
-
Improving Sparse Autoencoder with Dynamic Attention
A cross-attention SAE with sparsemax attention achieves lower reconstruction loss and higher-quality concepts than fixed-sparsity baselines by making activation counts data-dependent.
-
MetaSAEs: Joint Training with a Decomposability Penalty Produces More Atomic Sparse Autoencoder Latents
Joint training of a primary SAE with a meta SAE that applies a decomposability penalty on decoder directions produces more atomic latents, shown by 7.5% lower mean absolute phi and 7.6% higher fuzzing scores on GPT-2.
-
Scaling and evaluating sparse autoencoders
K-sparse autoencoders with dead-latent fixes produce clean scaling laws and better feature quality metrics that improve with size, shown by training a 16-million-latent model on GPT-4 activations.
-
Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
Sparse autoencoders scaled to 34 million features on Claude 3 Sonnet yield interpretable, steerable representations of concrete and abstract concepts that generalize across languages and modalities.
-
DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices
DECO is a sparse MoE architecture with ReLU-based routing, learnable expert scaling, and NormSiLU activation that matches dense Transformer performance at 20% expert activation and delivers 2.93x speedup on Jetson AGX Orin.
-
Emergent Symbolic Structure in Health Foundation Models: Extraction, Alignment, and Cross-Modal Transfer
Health foundation model embeddings contain an interpretable symbolic organization shared across modalities that supports cross-domain transfer without joint training.
-
Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders
Sparse autoencoders isolate unstable features in reward model representations and enable two mitigation techniques that reduce preference errors on perturbed inputs without retraining.
-
From Token Lists to Graph Motifs: Weisfeiler-Lehman Analysis of Sparse Autoencoder Features
Graph-motif clustering of SAE features via a frequency-binned WL kernel recovers structural families not captured by decoder cosine similarity or token histograms.
-
Feature Starvation as Geometric Instability in Sparse Autoencoders
Adaptive elastic net SAEs (AEN-SAEs) mitigate feature starvation in SAEs by combining ℓ2 structural stability with adaptive ℓ1 reweighting, producing a Lipschitz-continuous sparse coding map that recovers global feature support under mild assumptions.
-
Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs
DACO curates a 15,000-concept dictionary from 400K image-caption pairs and uses it to initialize an SAE that enables granular, concept-specific steering of MLLM activations, raising safety scores on MM-SafetyBench and JailBreakV while preserving general capabilities.
-
Sparse Autoencoders as a Steering Basis for Phase Synchronization in Graph-Based CFD Surrogates
Sparse autoencoders enable phase synchronization in frozen graph CFD surrogates through Hilbert-identified oscillatory features and SVD-based time-varying rotations.
-
MetaMorph: Multimodal Understanding and Generation via Instruction Tuning
VPiT enables pretrained LLMs to perform both visual understanding and generation by predicting discrete text tokens and continuous visual tokens, with understanding data proving more effective than generation-specific data.
-
Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
Sparse feature circuits are introduced as interpretable causal subnetworks in language models, supporting unsupervised discovery of thousands of circuits and a method called SHIFT to improve classifier generalization by ablating irrelevant features.
-
Steered Generation via Gradient-Based Optimization on Sparse Query Features
Prototype-Based Sparse Steering decomposes query activations with SAEs and optimizes sparse features via gradients to steer LLM outputs toward specific behaviors.
-
From Correlation to Cause: A Five-Stage Methodology for Feature Analysis in Transformer Language Models
A five-stage causal feature analysis methodology is proposed and tested on GPT-2 for IOI, showing partial causality of SAE features, robustness differences under shifts, and deployment cost benefits.
-
Towards Effective Theory of LLMs: A Representation Learning Approach
RET learns temporally consistent macrovariables from LLM activations via self-supervised learning to support interpretability, early behavioral prediction, and causal intervention.
-
Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models
The survey organizes mechanistic interpretability techniques into a Locate-Steer-Improve framework to enable actionable improvements in LLM alignment, capability, and efficiency.