Alignment-Enhanced Decoding:Defending via Token-Level Adaptive Refining of Probability Distributions | Synapse