arXiv:2607. 20652v1 Announce Type: cross Abstract: Language models are thought to exhibit the phenomenon of superposition, representing many more features than dimensions in their residual streams.
arXiv:2606. 26620v1 Announce Type: cross Abstract: Sparse autoencoders (SAEs) have emerged as a powerful tool for decomposing superposed language model representations into sparse and interpretable features.
arXiv:2606. 14990v1 Announce Type: cross Abstract: Sparse autoencoders (SAEs) are standard tools for mechanistic interpretability, but current SAE families are constrained by fixed encoder nonlinearities such as ReLU, JumpReLU, and TopK.
arXiv:2606. 11722v1 Announce Type: cross Abstract: Finding interpretable directions in language-model representations is critical for understanding and controlling model behavior.
arXiv:2606. 14040v1 Announce Type: new Abstract: Sparse autoencoders (SAEs) are typically trained to reconstruct the \textbf{entire} residual stream through a sparse dictionary, implicitly assuming that all activation content is amenable to sparse, monosemantic decomposition.
arXiv:2606. 07414v1 Announce Type: new Abstract: Sparsity allows scaling model parameters without proportionally increasing computational cost.
arXiv:2607. 21291v1 Announce Type: cross Abstract: Large language models (LLMs) achieve strong generation and reasoning performance, but the Transformer architecture incurs high inference cost.
Scaling Large Language Models (LLMs) has been driven mainly by enlarging the Transformer backbone, but for an already-strong model this requires another round of costly pretraining. We study whether an existing backbone can keep improving by allocating more computation to each token while leaving the Transformer backbone fixed.
arXiv:2606. 06333v1 Announce Type: new Abstract: Sparse Autoencoders (SAEs) are widely used for mechanistic interpretability in large language models, yet their formulation assigns each latent feature a single decoder direction, implicitly assuming features to be one-dimensional.
The paper introduces GaugeLasso, a method that applies symmetric group‑lasso penalties to transformer channels during training, enabling entire tensor slices to be zeroed out while maintaining dense tensors for GPU efficiency. By calibrating channel penalties based on inference utility per compute, the network self‑organizes into depth‑dependent structural profiles that can be dramatically smaller than the original architecture, achieving up to 255‑fold compression on a polynomial division task and outperforming hand‑designed baselines on language modeling and autoencoding benchmarks. The approach also accelerates training and reveals over‑provisioned axes that guide subsequent design iterations.
arXiv:2606. 16112v1 Announce Type: cross Abstract: Residual architectures are ubiquitous in deep learning, but they suffer from a subtle structural limitation: the norm of the residual stream can grow rapidly with depth.
The paper proposes Bypass Observation, a non‑intrusive layer‑wise readout architecture that attaches read‑only observation heads to selected Transformer layers without feeding their outputs back into the backbone. Three variants are explored: a shared language‑model head across layers, layer‑specific heads, and a layer‑ or step‑adaptive head. The authors provide a closed‑form overhead estimate (≈ V/(12d)) and discuss ways to reduce cost, while distinguishing bypass chain‑of‑thought from conventional chain‑of‑thought and outlining potential applications to looped and recurrent‑depth Transformers.
arXiv:2608. 08888v1 Announce Type: new Abstract: Autoregressive transformers compute along two axes: horizontally across generated tokens, and vertically through model depth.