Using simple tree-search techniques for LLM token sampling can give better results

its_just_andy@alien.top · 2 years ago

Using simple tree-search techniques for LLM token sampling can give better results

kilust@alien.top · 2 years ago

That’s really interesting! Maybe an experiment with the viterbi algorithm with the goal to maximize similarity with the context (or another score) could also improve results 🤔

Using simple tree-search techniques for LLM token sampling can give better results

Using simple tree-search techniques for LLM token sampling can give better results

Trying simple tree-search techniques for LLM token sampling