Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effects

Statistically Grounded Sparse-Feature Interventions for Activation-Space Control in Large Language Models

Convolution for Large Language Models

These results support depthwise convolution as a lightweight complement to self-attention for modeling short-range token interactions and suggest that the convolution makes repeated token IDs more sensitive to their immediate context.

Persistent Sparse Autoencoders: Learning Feature-Specific Timescales in Language Model Representations

Sparse autoencoders (SAEs) decompose language model activations into sparse features, yet these models traditionally encode each token independently, failing to expose information that persists across a sequence. We first show that temporal persistence can naturally emerge in standard SAE features: after a feature activates, the hidden state remains aligned with its direction, and past activations help reconstruct later hidden states. How long this lasts varies widely across features. We therefore introduce Persistent Sparse Autoencoders (Persistent SAEs), an extension of standard SAEs that learns a persistence coefficient for each feature, allowing the model to learn feature-specific timescales from reconstruction alone. Our experiments show that Persistent SAEs retain competitive reconstruction quality while learning a spectrum of timescales: short-timescale (fast) features stay locally interpretable, whereas long-timescale (slow) features accumulate information that identifies the current context. Moreover, we show in a prompt-injection monitoring case study that slow features preserve injection-related signals and remain causally effective over long contexts. These results suggest that Persistent SAEs offer new opportunities for interpreting and monitoring language models via persistent sparse features.

Sparse Autoencoders Reveal Structural and Family-level Features in BiRNA-BERT

Motivation RNA language models learn representations that support structure and function prediction, but which biological concepts their hidden states encode remains unclear. Sparse autoencoders (SAEs) decompose hidden states into interpretable features, yet have not been applied to RNA language models, where byte-pair tokenization breaks the one-token-one-nucleotide correspondence that nucleotide-level attribution assumes. Results We present SPIRAL, a layer-wise SAE analysis of BiRNA-BERT. Independent SAEs at layers 0, 5, and 11 expand each 768-dimensional hidden state into 6,144 features while preserving model behaviour (explained variance above 0.99997; masked-language-model sequence recovery near 99.7%). Tokenizer-aware offset propagation aligns features to nucleotides: at layer 5, 44.3% of tested features are significantly associated with bpRNA secondary-structure classes (mean enrichment 1.61 ×), and all 1,237 eligible features with RNAcentral RNA types. Sparse profiles raise k-nearest-neighbour balanced accuracy from 0.328 to 0.359 over dense embeddings at layer 5. Availability and Implementation Source code is available at https://github.com/SadatHossain01/SPIRAL; the code, evaluation data, and trained SAE checkpoints are archived at https://doi.org/10.5281/zenodo.21891845. Contact mrahman@cse.buet.ac.bd Supplementary information Supplementary data are presented alongside the manuscript.

Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects

The wide-scale use of sparse autoencoders (SAEs) as interpretability tools is limited by inconsistent links between SAE features and model behavior. Features with clear activation descriptions may have weak or unexpected causal effects; steering can vary across prompts or oppose the intended direction; and activation-based feature selection can miss features that produce the desired output change. Prior work has studied feature geometry inside the model, where features are computed. We instead study the geometry of changes in model logits caused by feature interventions. We introduce Feature-Effect Geometry Analysis (FEGA), an unsupervised framework that removes the same active SAE feature across contexts and analyzes the resulting cloud of logit changes. Across SAE variants, consistent one-dimensional effects are rare: few features behave like reusable directions. To interpret this variation, we distinguish value-like features, tied to static information such as factual attributes, from pointer-like features, associated with context-dependent operations. Value-like features more often exhibit structured, low-dimensional effects, although these effects typically span several directions. Pointer-like features, by contrast, predominantly exhibit diffuse effects. Our results show that a feature can be interpretable and causally relevant without providing a stable direction for steering.

Beyond a Bag of Features: Set-Level Instability in Sparse Autoencoders

It is found that SAE activation sets do not recover human category boundaries or within-category typicality more faithfully than dense embeddings or residual-stream states, but instead track model-internal similarity structure.