Anthropic's research uses effectiveness and helpfulness to characterize interference weights.

Anthropic researchers proposed analyzing the perturbation weights in neural networks based on their effectiveness and helpfulness, and observed that most high-impact weights are helpful to the model's behavior.

The experimental subject was a single-layer micro Transformer with approximately 2.9 million parameters, and the results cannot be directly extrapolated to current state-of-the-art large models.

refer to:Anthropic Transformer Circuits