The paper argues that the linear representation hypothesis—where neural activations encode concepts as linear mixtures—breaks down under superposition, causing linear decision boundaries to lose linearity after projection. Classical sparse coding with per‑sample inference can recover latent factors, but sparse autoencoders (SAEs) amortise this inference into a fixed encoder, creating a systematic gap that persists across training sizes, latent dimensions, and sparsity levels. Experiments show that the failure of SAEs under out‑of‑distribution compositional shifts stems from poor dictionary learning rather than inference, and that a good dictionary can solve the problem at all tested scales, highlighting scalable dictionary learning as the key open challenge.
By Vit\'oria Barin-Pacela, Shruti Joshi, Isabela Camacho, Simon Lacoste-Julien, David Klindt
The paper introduces Consistent Sparse Dictionary Learning (ConSDL), a method that replaces the linear encoder in sparse autoencoders with per‑sample sparse coding to improve dictionary learning. Synthetic experiments show ConSDL achieves near‑perfect recovery of dictionaries and coefficients, and in real vision and language tasks it boosts cross‑seed consistency by 17.6% for dictionaries and 35.9% for coefficients compared to six SAE baselines.
By Jinho Choi, Hyesu Lim, Jaegul Choo, Steffen Schneider
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.
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.
By Daniil Ognev, C\'elian Vasson, Lijie Hu, Kentaro Inui, Benjamin Heinzerling
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.
By Andrew Mack, Kraig Yuheng Tou, Mark Henry, Zhengxun Wu, Lauren Greenspan
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.
By Ankur Garg, Xuemin Yu, Hassan Sajjad, Samira Ebrahimi Kahou
arXiv:2607. 08605v1 Announce Type: cross Abstract: Sparse autoencoders (SAEs) have emerged as a promising technique for mechanistic interpretability by learning a set of sparse latent features in large models, each of which encodes a distinct concept.
By Weiduo Liao, Yunqiao Yang, Ying Wei
The paper introduces SCALPEL, a contrastive sparse autoencoder that selectively learns to unlearn target-specific information from neural network representations. It addresses the energy bias in standard reconstruction-based extractors, which tend to preserve dominant background structure while neglecting low-energy target components. The authors provide theoretical justification that contrastive training enhances target selectivity and demonstrate experimentally that SCALPEL outperforms NMF and standard SAE interventions on the TOFU benchmark, achieving performance comparable to Gradient Difference and RMU methods.
By Itai Zehavi, Fanny Jourdan, Ulrich Aivodji
arXiv:2606. 02385v1 Announce Type: cross Abstract: Sparse Autoencoders (SAEs) have found success parsing neural representations into interpretable concepts, providing a basis for understanding and control.
By William Dorrell
arXiv:2607. 17770v1 Announce Type: cross Abstract: Within Explainable Artificial Intelligence, mechanistic interpretability uses Sparse Autoencoders (SAEs) to extract more interpretable features from neural representations.
By Katarzyna Filus, Sebastian Pokuci\'nski
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.
By Haoran Jin, Xiting Wang, Shijie Ren, Hong Xie, Defu Lian
arXiv:2606.16193v2 Announce Type: replace-cross
Abstract: Multimodal Large Language Models (MLLMs) have demonstrated strong performance on vision-language tasks, yet their internal visual representat...
By Yusong Zhao, Hengyi Wang, Tanuja Ganu, Akshay Nambi, Hao Wang