Mechanistic interpretability of vision transformers seeks to decompose model computation into human-readable units, but learned representations entangle many concepts in each neuron. Feature superposi...
The paper investigates a spatial-concentration bias in Evolvable-Substrate HyperNEAT (ES‑HyperNEAT) when applied to MNIST, where evolved networks focus on a central cluster of input pixels. By partitioning the input image into 13 non‑overlapping spatial segments and evolving a separate expert network for each, the authors achieve a 43% mean accuracy—an 106% relative improvement over the baseline—without relying on data‑driven weighting. The study also introduces a receptive‑field diagnostic to detect silent input‑coverage collapse and a spatial‑partitioning remedy to restore full image coverage.
arXiv:2607. 16295v1 Announce Type: cross Abstract: Mechanistic interpretability has made significant strides in understanding neural network representations, with sparse dictionary learning (SDL) methods, most prominently sparse autoencoders, as a central paradigm.
arXiv:2607.01630v2 Announce Type: replace Abstract: Dynamic expansion methods for class-incremental learning (CIL) protect task-specific knowledge by growing dedicated tokens or subnetworks, yet our...
NeuronEye is a plug‑in framework that builds a sparse, concept‑level neuron vocabulary from intermediate vision‑language model (VLM) representations and selectively activates query‑relevant visual concepts during inference. It decomposes vision‑token states into an overcomplete sparse basis organized by concept clusters, uses the language query to activate relevant clusters, localizes the corresponding image patches, and injects the focused evidence back into the vision tokens, while a suppression mechanism attenuates dominant perceptual directions. Experiments on Qwen2.5‑VL‑7B and LLaVA‑1.6‑7B show that NeuronEye improves CV‑Bench overall accuracy by +3.1, boosts Distance by +9.5, and raises BLINK Multi‑view by +8.3, indicating that sparse neuron vocabularies can act as active interfaces for concept‑level visual reasoning.
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.
arXiv:2608. 14922v1 Announce Type: cross Abstract: Mechanistic interpretability has recently expanded to Vision Transformers (ViTs), with Sparse Autoencoders (SAEs) increasingly used as post-hoc tools to decompose internal representations into sparse and more interpretable features.
Dynamic expansion methods for class-incremental learning (CIL) protect task-specific knowledge by growing dedicated tokens or subnetworks, yet our analyses suggest that classification supervision alone does not sufficiently preserve task-agnostic shared backbone representations over long incremental sequences. We identify two intertwined challenges: cross-task confusion from sequential training on predominantly current-task data, which biases decision boundaries toward recent tasks; and under-optimized shared representations in the backbone that cap long-term discriminability as tasks accumulate.
arXiv:2608. 12408v1 Announce Type: cross Abstract: Representational similarity analysis (RSA) is increasingly used to ask which learning rules give convolutional networks brain-like representations.
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:2607. 15693v1 Announce Type: cross Abstract: We describe a model of perceptual inference in primary visual cortex (V1) equivalent to a minimal diffusion model whose function can be readily understood from its parameters.
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.