2 ms·
Does a LLM necessarily generate words and phrases with a similar distribution to human authors? Maybe people use “zipf’s law” but today’s LLMs can’t exactly mod
by joshuanapoli 3y ago
Does a LLM necessarily generate words and phrases with a similar distribution to human authors? Maybe people use “zipf’s law” but today’s LLMs can’t exactly model that. The result being some flattening of the distribution of generated phrases.
- espe 3y agointeresting question. anyone know of a paper that looks at power law distributions in llm vs. human text?