Injecting Random Tokens at the Decoding Stage Expands LLM Diversity

Harvard researchers proposed Recoding-Decoding (RD). At the decoding stage, it injects a random priming phrase and a short diverting stem to shake the LLM out of repeating only the most common answer path.

• It's designed for broadly exploring candidates in search quests without a fixed answer, such as wedding dresses, research topics, or startup names.
• It was evaluated on 50 brainstorming topics and 500 prompts from 5 public datasets, and relevance was maintained at around 0.99.
• As the number of repeated runs increased, the diversity and creativity of the outputs kept growing, and the effect was larger for stronger models.
• It was presented as a simple decoding method that can be applied without fine-tuning.