SAE High Frequency Latent

Group Bias Adaptation (GBA) adaptively adjusts neuron-specific biases to capture raw frequency characteristics by matching the Target Activation Frequency (TAF). For single-group BA, we mathematically proved that when data satisfies "sparse, non-cohesive, anti-coherence" conditions and neuron count, bias range, and learning rate requirements are met, all monosemantic features can be fully restored in O(1) iterations. As a result, we confirmed superior reconstruction loss, activation sparsity, and feature consistency compared to TopK and ℓ1 in LLMs with up to 1.5B parameters.