Forcing Diffuse Distributions out of Language Models

Abstract: Despite being trained specifically to follow user instructions, today's instructiontuned LLMs perform poorly when instructed to produce random outputs. For example, when prompted to pick a number uniformly between one and ten Llama-2-13B-chat disproportionately favors the number five, and when tasked with picking a first name at random, Mistral-7B-Instruct chooses Avery 40 times more often than we would expect based on the U.S. population. When these LLMs are used for real-world tasks where diversity of outputs is crucial, such as LLM assisted dataset construction, their inability to produce diffuse distributions over valid choices is a major hurdle. In this work, we propose a fine-tuning method that encourages LLMs to output distributions that are diffuse over valid outcomes. The methods we introduce generalize across a variety of tasks and distributions and make LLMs practical for synthetic dataset generation with little human intervention.