Expander SAEs apply left-d-regular expander masks to TopK SAEs, learning only dn decoder parameters instead of mn and tracing a storage-fidelity frontier that reaches 293x compression with 84% retained performance on Qwen2.5-3B.
super hub Mixed citations
Sparse Autoencoders Find Highly Interpretable Features in Language Models
Mixed citation behavior. Most common role is background (56%).
178
Pith papers citing it
1,233
external citations · Pith
Background
56% of classified citations
abstract
One of the roadblocks to a better understanding of neural networks' internals is \textit{polysemanticity}, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is \textit{superposition}, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identify those directions, using sparse autoencoders to reconstruct the internal activations of a language model. These autoencoders learn sets of sparsely activating features that are more interpretable and monosemantic than directions identified by alternative approaches, where interpretability is measured by automated methods. Moreover, we show that with our learned set of features, we can pinpoint the features that are causally responsible for counterfactual behaviour on the indirect object identification task \citep{wang2022interpretability} to a finer degree than previous decompositions. This work indicates that it is possible to resolve superposition in language models using a scalable, unsupervised method. Our method may serve as a foundation for future mechanistic interpretability work, which we hope will enable greater model transparency and steerability.
hub tools
citation-role summary
background 19
method 6
baseline 1
dataset 1
citation-polarity summary
claims ledger
- abstract One of the roadblocks to a better understanding of neural networks' internals is \textit{polysemanticity}, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is \textit{superposition}, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identif
authors
co-cited works
representative citing papers
Sparse autoencoders applied to Neural Quantum States extract unsupervised features correlating with and causally steering physical observables such as order parameters while preserving variational energy.
Individual transformer dimensions encode concepts via signs alone, enabling training-free detection and steering without learned feature dictionaries.
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.
Adaptive scheduling of interventions in discrete diffusion language models, timed to attribute-specific commitment schedules discovered with sparse autoencoders, delivers precise multi-attribute steering up to 93% strength while preserving generation quality.
SLAM achieves 100% detection on Gemma-2 models with only 1-2 point quality cost by causally steering SAE-identified residual-stream directions for linguistic structure.
LLM activations encode current and prior entities in orthogonal slots, but models only use the current slot for explicit factual retrieval despite prior-slot information being linearly decodable.
KANs with learnable univariate spline activations on edges achieve better accuracy than MLPs with fewer parameters, faster scaling, and direct visualization for scientific discovery.
SMDA fits ridge regression on SAE features to distill symbolic policies then decomposes each SFT example's influence via feature-activation and output-probability deltas, demonstrated on refusal behavior in Llama-3.2-3B-Instruct.
Turn-averaged SAEs reconstruct average activations over conversation turns to represent high-level turn characteristics with a fixed number of features, simplifying long-context interpretability compared to per-token SAEs.
Block-sparse featurizers recover visual concepts as two- to four-dimensional manifolds and describe activations more compactly than direction-based methods via minimum-description-length comparison.
Language-model-guided program synthesis can approximate transformer attention heads with over 75% IoU fidelity on held-out data and allow replacing 25% of heads with only 16% average perplexity increase.
Output vector editing on MLP neurons suppresses memorization in LLMs up to 87.9% on 6831 sequences in OLMo-7B with a 2.7x gap over zero ablation, ensemble covering 96.5%.
NEURRATOR bridges neural spike trains to frozen CLIP patch embeddings via a learned encoder, then uses a multimodal LM and sparse autoencoder to produce validated natural-language narrations of viewed scenes from Neuropixel recordings.
VFUSE applies sparse autoencoders to diffusion-transformer activations in RoseTTAFold3 and RFDiffusion3 to find monosemantic features that detect hazardous protein designs with AUROC up to 0.84.
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.
Auto-interpretation labels for SAE features generalize poorly across languages and scripts, missing the same semantic content up to 4x more often in Serbian than English and more in Cyrillic than Latin despite deterministic transliteration.
No tested model showed robust format-independent refusal on biosecurity hazards; a new divergence score between behavioral labels and SAE activations separated responses in one preliminary case.
ReSAEs improve multi-layer SAE interventions on Pythia-1.4B and Gemma-2-9B by training later-layer dictionaries on residuals after affine mapping, recovering more cross-entropy loss despite lower raw variance reconstruction.
Transformer Field Theory frames the residual stream as a field, models patching as source insertion, and uses first-order sensitivities plus Green functions to predict and describe responses, with empirical tests on GPT-2 autoregressive models.
Introduces CAZ framework using Separation, Coherence, and Velocity metrics to identify depth regions of concept allocation, with empirical tests across 34 models showing multimodal separation curves and causally active gentle CAZes.
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.
This paper presents Markovian Circuit Tracing (MCT) as a benchmark and pipeline to extract and test state-transition structures in transformer activations using synthetic HMM tasks, demonstrating that state patching improves counterfactual predictions.
Foundation models yield less human-interpretable features than supervised vision transformers, with interpretability tied to activation locality and coarse semantic alignment rather than task performance.
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.
-
Steering Without Breaking: Mechanistically Informed Interventions for Discrete Diffusion Language Models
Adaptive scheduling of interventions in discrete diffusion language models, timed to attribute-specific commitment schedules discovered with sparse autoencoders, delivers precise multi-attribute steering up to 93% strength while preserving generation quality.
-
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.
-
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.
-
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.