The paper introduces fundamental representational units (FRUs) in neural networks and proposes two evaluation criteria—fidelity and stability—to assess them. It demonstrates that threshold-activated sparse autoencoders (TSAEs) can identify FRUs with high fidelity and stability, supported by large-scale experiments. The study further shows that more stable FRUs are more monosemantic and can be combined across layers to reconstruct semantic structures that align with human conceptual organization.
arXiv:2606. 07007v1 Announce Type: cross Abstract: We propose a unified mathematical framework for a geometric understanding of concept learning and neuron interpretation in sparse autoencoders (SAEs).
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.
The paper investigates whether neural networks exhibit conceptual separation, meaning that examples of the same concept cluster together and related concepts are closer in representation space. Using geometric and distributional analyses, the authors find that Convolutional Neural Networks (CNNs) produce coherent, semantically ordered representations for familiar ImageNet concepts, but this coherence weakens for unseen concepts and under domain shift. Large Language Models (LLMs) keep distinct domains well separated, bring related subdomains closer, yet lose distinction between ambiguous topics at both mean and covariance levels.
arXiv:2608. 11197v1 Announce Type: new Abstract: Shani et al.
FlowGenesis is a framework that lets neural networks self‑organize their architecture during training by adding or removing neurons based on a local statistic called Representation Flux. Neurons with high Flux are split to preserve function, while low‑Flux, low‑activity neurons are pruned. Experiments on image and text classification show that networks starting from different sizes converge to similar neuron counts and performance, and the method also improves continual learning by reducing parameters for new tasks without sacrificing accuracy.
arXiv:2608.28184v1 Announce Type: new Abstract: Grokking is a delayed transition from memorization to generalization that is often accompanied by substantial reorganization of internal representation...
arXiv:2608.29034v1 Announce Type: cross Abstract: A wide range of methods have been proposed for interpreting language models, delivering important insights into their inner workings. However, differ...
arXiv:2608.29530v1 Announce Type: cross Abstract: Modern systems in artificial intelligence (AI) somehow excel in domains for which they seem poorly suited. Intelligence has traditionally been modele...
arXiv:2606. 07414v1 Announce Type: new Abstract: Sparsity allows scaling model parameters without proportionally increasing computational cost.
arXiv:2611.13270v1 Announce Type: new Abstract: Recent work has identified a set of emergent symbolic mechanisms that support abstract reasoning in large language models, but it remains unclear what...
arXiv:2606. 16193v1 Announce Type: cross Abstract: Multimodal Large Language Models (MLLMs) have demonstrated strong performance on vision-language tasks, yet their internal visual representations remain difficult to interpret.
arXiv:2608. 06839v1 Announce Type: new Abstract: Artificial Neural networks (ANNs) are often treated as black-box models, making explainability a central challenge in deep learning.