Nir N Shavit

Learning to Interpret Weight Differences in Language Models

Negative Pre-activations Differentiate Syntax

Budgeted Broadcast: An Activity-Dependent Pruning Rule for Neural Network Efficiency

Scalable Energy-Based Models via Adversarial Training: Unifying Discrimination and Generation

Expand Neurons, Not Parameters

Toy Models of Combinatorial Interpretability

On the Complexity of Neural Computation in Superposition

Wasserstein Distances, Neuronal Entanglement, and Sparsity