Activation Logging and Internal State Monitoring One foundational approach is activation logging, which involves recording the internal activations (neuron outputs, attention patterns, etc.) of a model during its forward pass. By inspecting these activations, researchers can identify which parts of the network are highly active or contributing to a given output. Many open-source transformer models […]
