Decoding-time Intervention

An intervention that influences the model's output probability distribution during autoregressive generation, such as logit bias, probability reweighting or constrained decoding, without modifying model parameters or input prompts.