We present a method for extracting \emph{monosemantic} neurons, defined as latent dimensions that align with coherent and interpretable concepts, from user and item embeddings in recommender systems. Our approach employs a Sparse Autoencoder (SAE) to reveal semantic structure within pretrained representations. In contrast to work on language models, monosemanticity in recommendation must preserve the interactions between separate user and item embeddings. To achieve this, we introduce a \emph{prediction aware} training objective that backpropagates through a frozen recommender and aligns the learned latent structure with the model's user-item affinity predictions. The resulting neurons capture properties such as genre, popularity, and temporal trends, and support post hoc control operations including targeted filtering and content promotion without modifying the base model. Our method generalizes across different recommendation models and datasets, providing a practical tool for interpretable and controllable personalization. Code and evaluation resources are available at https://github.com/DeltaLabTLV/Monosemanticity4Rec.

In the past 20 years, artificial neural networks have become dominant in various areas, continually growing in scale. However, the current analysis of large models has mainly focused on functionality, overlooking the influence of scale differences on their properties. To address this, we propose the concept of Emergence Learning, which emphasizes the significance of scale. By studying models of different scales, we have identified a key factor in achieving higher performance in large models: the decrease of monosemantic neurons. Building on this insight, we propose a proactive approach to inhibit monosemanticity for improved performance. Our solution involves a two-phase process that includes monosemantic neuron detection and inhibition, supported by theoretical analysis. Experimental results on various tasks and neural networks demonstrate the effectiveness of our proposed method. Following the idea of Emergence Learning, though drawing inspiration from scaling phenomena, the applicability of our method is not restricted to large scale alone. Therefore, the experiment is self-contained. However, extending this research to very large-scale datasets is appealing yet impossible for research departments due to limited resources. We are delighted to share the first co-authorship and eagerly await collaboration from any AI company before submission.

In some neural networks, individual neurons correspond to natural ``features'' in the input. Such \emph{monosemantic} neurons are of great help in interpretability studies, as they can be cleanly understood. In this work we report preliminary attempts to engineer monosemanticity in toy models. We find that models can be made more monosemantic without increasing the loss by just changing which local minimum the training process finds. More monosemantic loss minima have moderate negative biases, and we are able to use this fact to engineer highly monosemantic models. We are able to mechanistically interpret these models, including the residual polysemantic neurons, and uncover a simple yet surprising algorithm. Finally, we find that providing models with more neurons per layer makes the models more monosemantic, albeit at increased computational cost. These findings point to a number of new questions and avenues for engineering monosemanticity, which we intend to study these in future work.