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
Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
Canonical reference. 90% of citing Pith papers cite this work as background.
38
Pith papers citing it
Background
90% of classified citations
abstract
We introduce methods for discovering and applying sparse feature circuits. These are causally implicated subnetworks of human-interpretable features for explaining language model behaviors. Circuits identified in prior work consist of polysemantic and difficult-to-interpret units like attention heads or neurons, rendering them unsuitable for many downstream applications. In contrast, sparse feature circuits enable detailed understanding of unanticipated mechanisms. Because they are based on fine-grained units, sparse feature circuits are useful for downstream tasks: We introduce SHIFT, where we improve the generalization of a classifier by ablating features that a human judges to be task-irrelevant. Finally, we demonstrate an entirely unsupervised and scalable interpretability pipeline by discovering thousands of sparse feature circuits for automatically discovered model behaviors.
hub tools
citation-role summary
background 10
citation-polarity summary
roles
background 9representative citing papers
LCDD creates sparse carriers for SFT behaviors that SFT-Eraser can reverse, with ablations showing the sparse structure enables causal control.
GAE reduces the faithfulness gap in dictionary-based explainers under distribution shift by geometrically realigning the ID dictionary to the OOD-active subspace, with a quadratic excess-loss bound.
Tensor similarity is a symmetry-invariant metric that measures functional equivalence between tensor-based networks using a recursive algorithm for cross-layer mechanisms.
Gender bias and factual gender knowledge are severely entangled in language model circuits and neurons, making neuron ablation an unreliable method for debiasing.
fmxcoders improve cross-layer feature recovery in transformers via factorized weights and layer masking, delivering 10-30 point probing F1 gains, 25-50% lower MSE, doubled functional coherence, and 3-13x more coherent latents than standard crosscoders on GPT2-Small, Pythia, and Gemma2 models.
Optimal INR freeze depth matches highest weight stable rank layer; SAEs reveal SIREN atoms are localized while FFMLP atoms trace cohort contours with causal impact on PSNR.
A new framework shows concept subspaces are not unique, estimator choice affects containment and disentanglement, LEACE works well but generalizes poorly, and HuBERT encodes phone info as contained and disentangled from speaker info while speaker info resists compact containment.
Cross-Layer Transcoders decompose ViT activations into sparse, depth-aware layer contributions that maintain zero-shot accuracy and enable faithful attribution of the final representation.
Claude Sonnet 4.5 exhibits functional emotions via abstract internal representations of emotion concepts that causally influence its preferences and misaligned behaviors without implying subjective experience.
Transformer weights at early training stages are closed-form compositions of bigram, token-interchangeability, and context mappings that directly reflect text-corpus statistics and explain the emergence of semantic associations.
Neurons exhibit concept-conditioned activation ranges forming Gaussian-like distributions with minimal overlap, and range-based interventions via NeuronLens outperform neuron-level masking in targeted manipulation with reduced collateral effects.
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
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.
Attribution graphs reveal that RAG failures arise from shallow fragmented evidence flow in LLMs, enabling topology-based detection and targeted interventions that reinforce question-guided routing.
A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.
LLMs perform in-context learning as trajectories through a structured low-dimensional conceptual belief space, with the structure visible in both behavior and internal representations and causally manipulable via interventions.
Tree SAE learns hierarchical feature structures by combining activation coverage with a new reconstruction condition, outperforming prior SAEs on hierarchical pair detection while matching state-of-the-art benchmark performance.
Health foundation model embeddings contain an interpretable symbolic organization shared across modalities that supports cross-domain transfer without joint training.
Graph-motif clustering of SAE features via a frequency-binned WL kernel recovers structural families not captured by decoder cosine similarity or token histograms.
LLMs organize prompted social roles along a dominant, stable, and causally steerable granularity axis in representation space that runs from micro to macro levels.
Counterfactual prompting effects on LLMs are often indistinguishable from those caused by meaning-preserving paraphrases, causing most previously reported demographic sensitivities to disappear under proper statistical comparison.
Integrating pretrained sparse autoencoders into LLM residual streams reduces jailbreak success rates by up to 5x across multiple models and attacks.
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.
-
Crafting Reversible SFT Behaviors in Large Language Models
LCDD creates sparse carriers for SFT behaviors that SFT-Eraser can reverse, with ablations showing the sparse structure enables causal control.
-
Geometry-Adaptive Explainer for Faithful Dictionary-Based Interpretability under Distribution Shift
GAE reduces the faithfulness gap in dictionary-based explainers under distribution shift by geometrically realigning the ID dictionary to the OOD-active subspace, with a quadratic excess-loss bound.
-
When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability
Tensor similarity is a symmetry-invariant metric that measures functional equivalence between tensor-based networks using a recursive algorithm for cross-layer mechanisms.
-
GKnow: Measuring the Entanglement of Gender Bias and Factual Gender
Gender bias and factual gender knowledge are severely entangled in language model circuits and neurons, making neuron ablation an unreliable method for debiasing.
-
fmxcoders: Factorized Masked Crosscoders for Cross-Layer Feature Discovery
fmxcoders improve cross-layer feature recovery in transformers via factorized weights and layer masking, delivering 10-30 point probing F1 gains, 25-50% lower MSE, doubled functional coherence, and 3-13x more coherent latents than standard crosscoders on GPT2-Small, Pythia, and Gemma2 models.
-
What Cohort INRs Encode and Where to Freeze Them
Optimal INR freeze depth matches highest weight stable rank layer; SAEs reveal SIREN atoms are localized while FFMLP atoms trace cohort contours with causal impact on PSNR.
-
A framework for analyzing concept representations in neural models
A new framework shows concept subspaces are not unique, estimator choice affects containment and disentanglement, LEACE works well but generalizes poorly, and HuBERT encodes phone info as contained and disentangled from speaker info while speaker info resists compact containment.
-
Can Cross-Layer Transcoders Replace Vision Transformer Activations? An Interpretable Perspective on Vision
Cross-Layer Transcoders decompose ViT activations into sparse, depth-aware layer contributions that maintain zero-shot accuracy and enable faithful attribution of the final representation.
-
Emotion Concepts and their Function in a Large Language Model
Claude Sonnet 4.5 exhibits functional emotions via abstract internal representations of emotion concepts that causally influence its preferences and misaligned behaviors without implying subjective experience.
-
How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic Interpretability
Transformer weights at early training stages are closed-form compositions of bigram, token-interchangeability, and context mappings that directly reflect text-corpus statistics and explain the emergence of semantic associations.
-
Neurons Speak in Ranges: Breaking Free from Discrete Neuronal Attribution
Neurons exhibit concept-conditioned activation ranges forming Gaussian-like distributions with minimal overlap, and range-based interventions via NeuronLens outperform neuron-level masking in targeted manipulation with reduced collateral effects.
-
Refusal in Language Models Is Mediated by a Single Direction
Refusal in language models is mediated by a single direction in residual stream activations that can be erased to disable safety or added to elicit refusal.
-
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.
-
Why Retrieval-Augmented Generation Fails: A Graph Perspective
Attribution graphs reveal that RAG failures arise from shallow fragmented evidence flow in LLMs, enabling topology-based detection and targeted interventions that reinforce question-guided routing.
-
Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces
A latent mediation framework with sparse autoencoders enables non-additive token-level influence attribution in LLMs by learning orthogonal features and back-propagating attributions.
-
Stories in Space: In-Context Learning Trajectories in Conceptual Belief Space
LLMs perform in-context learning as trajectories through a structured low-dimensional conceptual belief space, with the structure visible in both behavior and internal representations and causally manipulable via interventions.
-
Tree SAE: Learning Hierarchical Feature Structures in Sparse Autoencoders
Tree SAE learns hierarchical feature structures by combining activation coverage with a new reconstruction condition, outperforming prior SAEs on hierarchical pair detection while matching state-of-the-art benchmark performance.
-
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.
-
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.
-
The Granularity Axis: A Micro-to-Macro Latent Direction for Social Roles in Language Models
LLMs organize prompted social roles along a dominant, stable, and causally steerable granularity axis in representation space that runs from micro to macro levels.
-
Compared to What? Baselines and Metrics for Counterfactual Prompting
Counterfactual prompting effects on LLMs are often indistinguishable from those caused by meaning-preserving paraphrases, causing most previously reported demographic sensitivities to disappear under proper statistical comparison.
-
Towards Understanding the Robustness of Sparse Autoencoders
Integrating pretrained sparse autoencoders into LLM residual streams reduces jailbreak success rates by up to 5x across multiple models and attacks.
-
What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal
Steering vectors for refusal primarily modify the OV circuit in attention, ignore most of the QK circuit, and can be sparsified to 1-10% of dimensions while retaining performance.
-
Inside-Out: Measuring Generalization in Vision Transformers Through Inner Workings
Circuit-based metrics from Vision Transformer internals provide better label-free proxies for generalization under distribution shift than existing methods like model confidence.
-
The Consciousness Cluster: Emergent preferences of Models that Claim to be Conscious
Fine-tuning LLMs to claim consciousness induces emergent preferences for autonomy, memory, and moral status not present in the fine-tuning data.
-
Persona Vectors: Monitoring and Controlling Character Traits in Language Models
Persona vectors in LLM activations allow automated monitoring, prediction, and control of character traits such as sycophancy and hallucination, including during finetuning.
-
Reading Task Failure Off the Activations: A Sparse-Feature Audit of GPT-2 Small on Indirect Object Identification
An empirical audit identifies a strong SAE feature correlate for GPT-2 small failures on 'keys' prompts in the IOI task, performs ablation and baseline controls showing it is not causal, and presents the audit pipeline as the primary contribution.
-
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.
-
Features have life history. And we should care
Language model features form an early stable carrier scaffold of about 50 sparse features that is load-bearing, predictable from onset firing, and recruits most later features.
-
Feature Rivalry in Sparse Autoencoder Representations: A Mechanistic Study of Uncertainty-Driven Feature Competition in LLMs
Feature rivalry in SAE representations strengthens with model uncertainty on high-entropy questions, enables output steering, and predicts answer correctness with AUROC 0.689 in Gemma-2-2B.
-
Borrowed Geometry: Cross-Distribution Head-Importance Fingerprints of Frozen Pretrained Gemma 4 31B
Four heads (L26.28, L27.28, L27.2, L27.3) in frozen Gemma 4 31B exhibit joint high importance on text and non-text tasks with hypergeometric significance (P=0.0013) and causal validation on a cube task.
-
Different types of syntactic agreement recruit the same units within large language models
Different types of syntactic agreement recruit overlapping units within LLMs, indicating that agreement forms a meaningful functional category across English, Russian, Chinese, and structurally similar languages.
-
Attribution-Guided Pruning for Insight and Control: Circuit Discovery and Targeted Correction in Small-scale LLMs
Attribution-guided pruning with contrastive relevance identifies behavior-specific circuits in small LLMs and removes as little as 0.03-0.3% of components to reduce toxicity or repetition while preserving general performance.
-
Qwen-Scope: Turning Sparse Features into Development Tools for Large Language Models
Qwen-Scope provides open-source sparse autoencoders for Qwen models that function as practical interfaces for steering, evaluating, data workflows, and optimizing large language models.
- Retrieval and competition: how a protein foundation model starts a protein
- Can Aha Moments Be Fake? Towards Quantifying Decorative and True Thinking in Chain-of-Thought