Seonglae ChoSelf Logits Evolution Decoding
A new decoding method that improves factuality by utilizing logit information from all layers. It reduces hallucinations by aligning the model's internal knowledge without external data or additional fine-tuning. SLED reuses logits from all layers to obtain probability distributions for each, then combines them via weighted average to select more accurate tokens.
Making LLMs more accurate by using all of their layers
Cyrus Rashtchian, Research Scientist, and Da-Cheng Juan, Research Lead, Google Research
https://research.google/blog/making-llms-more-accurate-by-using-all-of-their-layers/
