arXiv:2606. 06664v1 Announce Type: cross Abstract: Despite high accuracy, Vision Transformer (ViT) predictions can be driven by spurious cues, raising the need to understand their inner workings before safe deployment.
arXiv:2607. 07316v1 Announce Type: new Abstract: This article offers a comprehensive overview of mechanistic interpretability, an emerging field that seeks to reverse-engineer the internal algorithms of modern neural networks.
The paper introduces Polytopal Neural Networks (PNNs), a framework that enforces a polytope-based structure on neural network representations to extract distinct layer-wise aspects. By using learned corpus representations and an amortized simplex inference procedure, PNNs preserve meaningful latent structures with minimal performance loss and yield compressed representations comparable to vector‑quantized (VQ) methods. The authors demonstrate that PNNs provide a new, interpretable approach to VQ deep learning training while maintaining competitive performance.
The paper presents a detailed study of feature representations in Vision Transformers (ViTs), revealing that feature collapse at initialization causes redundancy and proposing a reduction scheme to address it. It demonstrates that entropy and minimum eigenvalue metrics reliably predict generalization, and that token‑space features and linear submodule outputs are particularly informative. The authors develop a proxy that improves ranking of ViT architectures by 18‑48% over previous baselines, enabling the selection of models with higher accuracy at comparable or lower computational cost.
arXiv:2606. 07414v1 Announce Type: new Abstract: Sparsity allows scaling model parameters without proportionally increasing computational cost.
arXiv:2607. 00089v1 Announce Type: new Abstract: Mechanistic interpretability has produced a rich inventory of component-level analyses that characterise what neural-network components encode and how they interact.
arXiv:2606. 11722v1 Announce Type: cross Abstract: Finding interpretable directions in language-model representations is critical for understanding and controlling model behavior.
The paper compares latent representations in Selective State Space Models (SSMs) like Mamba and Transformers such as Pythia using Sparse Autoencoders. Across a 10‑million token corpus, 99.98% of Mamba features align closely with Pythia’s, supporting the Universality Hypothesis that core semantic representations are similar across architectures. A tiny 0.02% of features diverge, with Mamba’s recurrent bottleneck causing it to compress syntactic anomalies into polysemantic neurons, whereas Pythia’s attention can isolate distinct formatting edge‑cases.
arXiv:2505.09716v4 Announce Type: replace Abstract: Out-of-distribution (OOD) generalisation through composition requires a system to discover invariant properties from input-output associations and...
arXiv:2608.30720v1 Announce Type: new Abstract: Representational similarity is foundational to analyses of deep networks, yet distances between point-valued representations are not intrinsically tied...
The paper introduces ClusterMixer, a token mixer based on hierarchical clustering in hyperbolic space, designed to be transparent and interpretable. It forms the core of a new vision backbone called HCFormer, which incorporates multiple clustering strategies to maintain strong performance. Experiments show HCFormer surpasses existing backbones on tasks such as image classification, object detection, instance segmentation, and semantic segmentation.
arXiv:2601. 21944v3 Announce Type: replace Abstract: The widespread adoption of deep learning models in computer vision has intensified concerns about interpretability.