SharedSAE demonstrates that a single sparse autoencoder can replace multiple model‑specific SAEs by using a shared dictionary with model‑specific encoder‑decoder pairs. It preserves activation magnitudes, normalizes only selection scores, and supports single‑model inference via model dropout. Trained on four 1B‑scale language models, SharedSAE retains 96.6% of the mean explained variance of dedicated SAEs, shows higher cross‑model latent correlations, and allows efficient adaptation of new models to the shared latent space.
arXiv:2606. 30609v1 Announce Type: cross Abstract: Sparse Autoencoders (SAEs) are widely used to interpret large language models by decomposing activations into sparse, human-understandable features, but scaling to large dictionaries exposes fundamental challenges.
arXiv:2608. 09521v1 Announce Type: new Abstract: Activation-based tools are usually tied to one model's native hidden space, requiring probes, sparse autoencoders, and natural-language interpreters to be rebuilt or rediscovered for each new language model.
Sparse autoencoders (SAEs) are proposed to extract numerous features from large language model (LLM) representations, yet explaining these features still relies primarily on external observation. This reliance leads to superficial explanations inferred from observed model behavior and computational inefficiency from collecting such behavioral evidence at scale.
arXiv:2607. 17117v1 Announce Type: new Abstract: Sparse autoencoders (SAEs) decompose language model activations into sparse features, but standard SAEs encode each token independently and do not expose information that persists across a sequence.
arXiv:2506. 20040v3 Announce Type: replace-cross Abstract: Interpreting language models remains challenging due to the existence of residual stream, which linearly mixes and duplicates features across adjacent layers, causing single-layer analyses to miss this cross-layer structure.
arXiv:2604. 03532v2 Announce Type: replace-cross Abstract: Large language models (LLMs) show strong multilingual capabilities, yet reliably controlling the language of their outputs remains difficult.
We present a Procrustes-conditioned Joint End-to-end Top-K Sparse Autoencoder (SAE) for extracting cross-seed universal features from independently trained BERT models. Cross-seed feature universality is a fundamental challenge in mechanistic interpretability: because dictionary learning is non-convex, independently trained networks learn misaligned feature spaces, so apparently identical features may differ by random initialization.
arXiv:2607. 15232v1 Announce Type: cross Abstract: A tokenizer fixed at the start of pre-training allocates vocabulary in proportion to the pre-training corpus, reflecting the deployment priorities at that time.
The study compares an English-only and a bilingual decoder-only model, each 310 M parameters, trained on eight diverse languages while controlling for English exposure, compute, and document overlap. After aligning on shared English vocabulary, the authors find that token embeddings appear similar, but the deeper hidden states used for prediction diverge across models. This hidden‑state mismatch grows through middle transformer layers and persists despite controls, indicating that contextual processing differs between the models. "whyItMatters":"The findings show that embedding alignment can conceal significant internal representation differences, which is crucial for any downstream work that assumes aligned multilingual models are interchangeable."
Persistent Sparse Autoencoders (Persistent SAEs) extend standard sparse autoencoders by learning a persistence coefficient for each feature, enabling the model to capture feature‑specific timescales from reconstruction alone. The study shows that these persistent features maintain competitive reconstruction quality while distinguishing between short‑timescale, locally interpretable features and long‑timescale, context‑accumulating features. In a prompt‑injection monitoring case study, slow features were found to preserve injection‑related signals and remain causally effective over long contexts.
Sparse autoencoders (SAEs) are commonly used to interpret large language models, but their reliability after pruning is unclear. This study shows that pruning’s effect on an SAE is governed by perturbation energy, a covariance-weighted norm, and that magnitude pruning distorts the representation space by ignoring activation geometry. Activation-aware pruning methods such as Wanda and SparseGPT better preserve SAE behavior, and the authors find that middle layers are especially vulnerable, leading them to propose a layer‑wise sparsity allocation that reduces perplexity for a given sparsity level.
arXiv:2607. 04593v1 Announce Type: cross Abstract: Vision-Language Models (VLMs) have demonstrated impressive capabilities across different tasks, but their computational cost is dominated by the large number of visual tokens fed to the language model.