rational-activations

Introduces Rational Sparse Autoencoder (RSAE), which replaces fixed encoder activations with trainable rational functions, improving reconstruction and sparsity trade-offs on residual-stream activations of open-weight language models across multiple baseline families.