SLED: Self Logits Evolution Decoding for Improving Factuality in Large Language Models

From an optimization perspective, our SLED framework leverages the latent knowledge embedded within the LLM by contrasting the output logits from the final layer with those from early layers.