Decoding-steering methods

Modify output logits during generation rather than hidden states. Each processor wraps an auxiliary guide model.

Method	Guide models	Mechanism
ContrastiveDecoding	1 guide	target - alpha * guide
ProxyTuning	2 guides (tuned + base)	target + scale * (tuned - base)
SafeDecoding	1 expert	Blend first-m tokens
Nudging	1 safe model	Hard-switch when uncertain